Blog & News

Part 2

In the first part of 'C for Cycle Time', we explored the essence of cycle time in front-end wafer fabs and its significance for semiconductor companies. We introduced the operating curve, which illustrates the relationship between fab cycle time and factory utilization, as well as the power of predictability and the ripple effects cycle time can have across the supply chain. 

In part 2, we will explore strategies to enhance cycle time through advanced scheduling solutions, contrasting them with traditional methods. We will use the operating curve, this time to demonstrate how advanced scheduling and operational factors, such as product mix and factory load, can significantly impact fab cycle time. 

How wafer fabs can improve cycle time 

By embracing the principles of traditional Lean Manufacturing, essentially focused on reducing waste in production, cycle time can be effectively reduced ^[1]. Here are a few strategies that can help improve fab cycle time: 

• Improving maintenance strategies, for example moving from reactive to more proactive maintenance can improve cycle time with fewer breakdowns and more predictable tool availability ^[2]. 
• As noted in part 1, minimizing wasted time in batch formation and reducing the frequency of rework due to defects improves cycle time.  
• Purchasing faster tools. Although, this can be a time-consuming and costly undertaking. In-facility expansion may take up to a year, while the commencement of a new facility could extend to three years ^[3].
• Establishing optimal batching in diffusion poses a considerable challenge, given the intricate process constraints within the diffusion area, such as timelinks, as we’ve explained in a recent blog.
• Balancing cycle time of hot lots with average fab cycle time. Fabs often assign higher priority to hot lots, which can negatively impact the average cycle time of production lots ^[4].
• Developing the skills of existing operators and expediting the onboarding process for new operators could be another means of reducing variability in production, thus impacting cycle time.
  ‍

The implementation of an advanced AI scheduler can facilitate most of the strategies noted above, leading to an improvement in cycle time with significantly less effort demanded from a wafer fab compared to alternatives such as acquiring new tools. In the next sections we are going to see how this technology can make your existing tools move wafers faster without changing any hardware!

Applying an advanced AI scheduler to improve cycle time 

In this section, we delve into how an advanced AI scheduler (AI Scheduler) can maintain factory utilization while reducing cycle time. 

First let’s define what an AI Scheduler is. It is an essential fab software that has a core engine powered by AI models such as mathematical optimization. It possesses the ability to adapt to ongoing real-time changes in fab conditions, including variations in product mixes, tool downtimes, and processing times. Its output decisions can achieve superior fab objectives, such as improved cycle time, surpassing the capabilities of heuristic-based legacy scheduling systems. More aspects of an advanced AI scheduler can be found in our previous article, A is for AI. The AI Scheduler optimally schedules fab production in alignment with lean manufacturing principles. It achieves this by optimally sequencing lots and strategically batching and assigning them to tools. 

Figure 5 shows an example of how an AI Scheduler can successfully shift the cycle time from the original operating curve closer to the theoretical operating curve. As a result, cycle time is now 30 days at 60% factory utilization. This can be accomplished by enhancing fab efficiency through measures such as minimizing idle times, reducing re-work, and mitigating variability in operations, among other strategies. In the next sections, we will show two examples in metrology and diffusion how cycle time is improved with optimal scheduling. 

Reducing queuing times and tool utilization variability in metrology 

Many wafer fabs employ a tool pull-system for dispatching. In this approach, operators typically decide which idle tool to attend to, either based on their experience or at times, randomly. Once at the tool, they then select the highest priority lots from those available for processing. A drawback of this system is that operators don't have a comprehensive view of the compatibility between the lots awaiting processing, those in transit to the rack, and the tools available. This limited perspective can lead to longer queuing times and underutilized tools, evident in Figure 6.

An AI Scheduler addresses these inefficiencies. By offering an optimized workflow, it not only shortens the total cycle time but also minimizes variability in tool utilization. This in turn indirectly improves the cycle time of the toolset and overall fab efficiency. For example, Seagate deployed an AI Scheduler to photolithography and metrology bottleneck toolsets that were impacting cycle time. The scheduler reduced queue time by 4.3% and improved throughput by 9.4% at the photolithography toolset ^[5]. In the metrology toolset, the AI Scheduler reduced variability in tool utilization by 75% which resulted in reduced cycle time too, see Figure 7 ^[6].

‍

‍

Improving cycle time and optimal batching in diffusion

Diffusion is a toolset that poses operational complexities due to its intricate batching options and several coupled process steps between cleaning and various furnace operations ^[7]. Implementing an AI Scheduler can mitigate many of these challenges, leading to reduced cycle time:

1. Strategic Batching can reduce total cycle time in diffusion. To maximize the benefit of an AI Scheduler, the fab should provide good quality data.
2. Automated Furnace Loading: Typically, diffusion loading is accomplished via a pull-system from the furnace. This means that operators would revisit the cleaning area to manually pick the best batches, based on upcoming furnace availability. This approach often demands substantial resources and time, thereby increasing cost or cycle time. The AI Scheduler curtails this time considerably, freeing up operators for other essential tasks, which indirectly may reduce cycle time elsewhere.
3. Reduction of Timelinks Violations: A recent pilot implementation of an AI Scheduler in diffusion at a Renesas fab underscored its effectiveness. As displayed in Figure 8, timelink violations were significantly reduced. This minimizes the necessity for rework, further cutting down the cycle time, as explained earlier in the article.
   ‍

Maximizing the value of the AI Scheduler by integrating with other applications

In the above examples of photo, metrology and diffusion toolsets, the AI Scheduler can support operators to achieve consistently high performance. To enhance the efficiency of the scheduling system in fabs predominantly run by operators with minimal AMHS (Automated Material Handling Systems) presence, pairing the scheduler with an operator guidance application, as detailed in one of our recent blogs on user-focused digitalisation, can be a valuable approach. This software will suggest the next task required to be executed by an operator. 

The deployment of an AI Scheduler should focus on bottleneck toolsets - specifically, those that determine the fab's cycle time. Reducing the cycle time of a toolset will be inconsequential if that toolset is not a bottleneck. Consequently, fabs should consider the following two approaches:

1. Ensure the deployment of the AI Scheduler on the most critical toolsets to effectively address dynamic bottlenecks. This ensures that as bottlenecks shift, the AI Scheduler can promptly reduce the cycle time of the newly identified bottlenecked toolset. By doing so, fabs can consistently maintain a low cycle time.
2. The introduction of a global (or a fab wide) application layer – such as a solution that looks across all the toolsets and all lots across the whole line – can help coordinate all deployed AI Schedulers. This application should indicate which toolsets are bottlenecks and it should also adjust lot priorities or production targets per toolset to ensure a smooth flow across the line. The interaction between global applications and local scheduling applications can be seen in recent papers ^{[9] [10]}. 

Dealing with dynamic changes in the fab and understand trade-offs between competing objectives

Another factor to consider is that the actual operating curve of the fab is moving constantly based on changes in the operating conditions of the fab. For example, if the product mix changes substantially, this may impact the recipe distribution enabled in each tool and subsequently, the fab cycle time vs factory utilization curve would shift. The operating curve can also change if the fab layout changes, for example when new tools are added.

In Figure 9, we show an example wherein the cycle time versus factory utilization curve for product mix A shifts upward. This signifies an increased cycle time in the fab due to the recent changes in the product mix (and the factory utilization was slightly reduced under these new conditions). An autonomous AI Scheduler, as described by Sebastian Steele in a recent blog, should be able to understand the different trade-offs. For example, in Figure 10, the AI Scheduler could deal with the same utilization as before (60%) with product mix A, but the cycle time will stay at 50 days (10 days more than in the case with product mix A). Another alternative is that the user can then decide if they want to customize this trade-off so that the fab can move back to the same cycle time with this new product mix B at 40 days but staying with lower utilization at 57%. 

Trade-offs between different objectives at local toolsets may impact the fab cycle time. Consider the trade-offs in terms of batching costs versus cycle time. For instance, constructing larger batches might be crucial for high-cost operational tools such as furnaces in diffusion and implant. However, this approach could lead to an extended cycle time for the specific toolset and, consequently, an overall increase in fab cycle time. 

Tool availability and efficiency significantly affect cycle time, akin to the influence of product mix on operating curves. If tools experience reduced reliability over time, the operating curve may shift upward, resulting in a worse cycle time for the same utilization. While the scheduler cannot directly control tool availability, strategically scheduling maintenance and integrating it with lot scheduling can positively impact cycle time. A dedicated future article will delve into this topic in more detail.

Conclusion

The topic of the cycle time has been enriched with the introduction of an AI Scheduler, bringing a paradigm shift in how we perceive and manage the dynamics of front-end wafer fabs. As highlighted in our exploration, these schedulers do more than just automate – they optimize. By understanding and predicting the nuances of operations, from tool utilization to lot prioritization, advanced AI schedulers provide a roadmap to not just manage but optimize cycle time considering alternative trade-offs. In future articles we will talk about how scheduling maintenance and other operational aspects can be considered in a unified and autonomous AI platform that we believe would be the next revolution, after the innovations from Arsenal of Venice, Ford and Toyota. 

‍

Author: Dennis Xenos, CTO and Cofounder, Flexciton

References

• [1] James P. Ignizio, 2009, Optimizing Factory Performance: Cost-Effective Ways to Achieve Significant and Sustainable Improvement 1st Edition, McGraw-Hill, ISBN 978-0-07-163285-0
• [2] Lean Production, 2023, TPM (Total Productive Maintenance), URL.
• [3] Ondrej Burkacky, Marc de Jong, and Julia Dragon, 2022, Strategies to lead in the semiconductor world, McKinsey Article, URL. 
• [4] Philipp Neuner, Stefan Haeussler, Julian Fodor, and Gregor Blossey, 2023, Putting a Price Tag on Hot Lots and Expediting in Semiconductor Manufacturing. In Proceedings of the Winter Simulation Conference (WSC '22). IEEE Press, 3338–3348.
• [5] Robert Moss, Dennis Xenos, Tina O’Donnell, 2023, Deployment of an Advanced Photolithography Scheduler at Seagate Technology, IFORS News, Volume 18, Issue 1, ISSN 2223-4373, pp. 8–10, URL.
• [6] Robert Moss, 2022, Ever-decreasing circles: how iterative modelling led to better performance at Seagate Technologies. Euro 2022 Conference, Finland, URL
• [7] Thomas Beeg, 2023, Impact of “time links” or controlled queue times, Factory Physics and Automation, URL.
• [8] Jamie Potter, 2023, Fab scheduling is now so complex that it needs next-generation intelligent software, Silicon Semiconductor Magazine, Volume 44, Issue 2, pp. 26-29, URL.
• [9] I. Konstantelos et al., 2022, "Fab-Wide Scheduling of Semiconductor Plants: A Large-Scale Industrial Deployment Case Study," 2022 Winter Simulation Conference (WSC), Singapore, pp. 3297-3308, doi: 10.1109/WSC57314.2022.10015364.
• [10] Félicien Barhebwa-Mushamuka. 2020,  Novel optimization approaches for global fab scheduling in semiconductor manufacturing. Other. Université de Lyon. English. ⟨NNT : 2020LYSEM020⟩. ⟨tel-03358300⟩

This two-part article aims to explain how we can improve cycle time in front-end semiconductor manufacturing through innovative solutions, moving beyond conventional lean manufacturing approaches. In part 1, we will discuss the importance of cycle time for semiconductor manufacturers and introduce the operating curve to relate cycle time to factory utilization. Part 2 will then explore strategies to enhance cycle time through advanced scheduling solutions, contrasting them with traditional methods.

Part 1 

Why manufacturers care about cycle time 

Cycle time, the time to complete and ship products, is crucial for manufacturers. James P. Ignazio, in Optimizing Factory Performance, noted that top-tier manufacturers like Ford and Toyota have historically pursued the same goal to outpace competitors: speed ^[1]. This speed is achieved through fast factory cycle times.

This emphasis on speed had tangible benefits: Ford, for instance, could afford to pay workers double the average wage while dominating the automotive market. The Arsenal of Venice's accelerated ship assembly secured its status as a dominant city-state. Similarly, fast factory cycle times were central to Toyota’s successful lean manufacturing approach.

Furthermore, semiconductor manufacturers grapple with extended cycle times that can often span 24 weeks ^[2]. This article will focus on manufacturing processes in front-end wafer fabs as their contribution to the end product, such as a chip or hard drive disk head, spans several months. In contrast, back-end processes can be completed in a matter of weeks ^[3]. However, the principles discussed apply universally to back-end fabs without sacrificing generality.

Why Short Cycle Times Matter for Front-end Wafer Fabs

• Revenue acceleration: The quicker products reach customers, the faster revenue streams in. However, quantifying the precise financial impact due to cycle time is intricate and beyond this article's scope.

• Competitive advantage: Gaining a competitive advantage involves reducing cycle time in R&D wafers, which accelerates product launches. More than 20% of front-end fab production lines can be used for R&D wafer testing and iteration. Swift deliveries enhance a company's reputation, leading to more contracts. At the 2022 Winter Simulation Conference, Micron highlighted their rapid advancements: maturing 30% quicker in DRAM (five months ahead of the previous node) and 20% in NAND (a year faster than the prior node). See Figure 1.

• Agility in market responsiveness: A fab with shorter cycle times can swiftly adjust to market fluctuations, whether that is a surge in demand or a shift in product preferences, such as changes in product mix. It can also respond faster to changes in customer requirements.

• Risk mitigation: The shorter the cycle time, the quicker a fab can respond once defects have been detected as it takes less time to perform rework.

• Inventory management: Lower cycle times can reduce the amount of work-in-progress (WIP) in buffers or racks (intermediate stock), or stock at the end of the production line. This not only liberates tied-up capital but also wafers can move quicker with less WIP in the fab as it is shown in a later section introducing the operating curve.
  ‍

Achieving Predictable Cycle Time

Less variability in cycle time helps a wafer fab to achieve better predictability in the manufacturing process. Predictability enables optimal resource allocation; for instance, operators can be positioned at fab toolsets (known as workstations) based on anticipated workload from cycle time predictions. Recognizing idle periods of tools allows for improved maintenance scheduling which will result in reduction in unplanned maintenance. In an upcoming article (Part 2), we'll explore how synchronizing maintenance with production can further shorten cycle times.

Measuring and monitoring the cycle time improves overall fab performance

Measuring and monitoring cycle times aids in identifying deviations from an expected variability. This, in turn, promptly highlights underlying operational issues, facilitating quicker issue resolution. Additionally, it assists industrial engineers in pinpointing bottlenecks, enabling a focused analysis of root causes and prompt corrective actions.

Supply chain stakeholders usually fail to understand the impact of cycle time 

In the semiconductor industry, cycle time plays a pivotal role in broader supply chain orchestration:

• A predictable cycle time informs suppliers when to provide fresh batches of raw materials. 
• Furthermore, it influences the downstream processes of Assembly & Test Operations (back-end facilities). Back-end facilities with a cycle time of less than a week gain enhanced predictability, allowing for more effective allocation of capacity and resources.
• Predictable cycle times will also inform safety inventory levels, freeing capital and optimizing storage space. 
  ‍

Cycle time is a component of the total lead time of a product (it also includes procurement, transportation, etc). Therefore, total lead time can be reduced if the long cycle times in the front-end wafer fabs are reduced. A reliable cycle time nurtures trust with suppliers, laying the foundation for favorable partnerships and agreements. In essence, cycle time is not just about production; it's the heartbeat of the semiconductor supply chain ecosystem.

Understanding how cycle time impacts product delivery times is essential for the semiconductor industry. In some analyses, you could see that cycle time is confused with capacity, as the authors in a McKinsey article stated “Even with fabs operating at full capacity, they have not been able to meet demand, resulting in product lead times of six months or longer” ^[4]. On the contrary, in a fab operating at full capacity, lead times of the products will increase as the average cycle time of manufacturing is increasing. 

How to measure cycle time

Fab Cycle Time

The fab cycle time metric defines the time required to produce a finished product in a wafer fab. The general cycle time term is also used to measure the time required to complete a specific process step (e.g. etching, coating) in a toolset, known as process step cycle time. The fab cycle time consists of the following time components as can be seen in Figure 2:

• Value-added processing time, which is the time taken to transform or assemble the unfinished product, which is a wafer in our case.

• Non-value-added processing time includes the time taken for inspection and testing, as well as the time for transferring the wafer between different steps.

• Time to prepare the products for processing: this refers to the time operators or tools required to form a batch, i.e. to select which lots should be bundled together for processing.

• Queue time: the time spent where the unfinished wafer is waiting to be processed because the tool required is busy, due to the tool processing another batch or undergoing maintenance.
  ‍

To measure and monitor cycle time, wafer fabs must track transactional data for each lot, capturing timestamps for events like the beginning and completion of processing at a tool. This data is gathered and stored by a Manufacturing Execution System (MES). Such transactional information can be utilized for historical operations analysis or for constructing models to forecast cycle times influenced by different operational factors. This foundation is crucial for formulating the operational curve of the fab, which we'll delve into in the subsequent part of this blog. As outlined in an article by Deenen et al., there are methods to develop data-driven simulations that accurately predict future cycle times ^[3].

Fab operating curve: Fab Cycle Time versus Factory Utilization

As we mentioned earlier, historic data can be used to generate the operating curve of a fab which describes the cycle time in relation to the factory utilization. Figure 3 shows the graph of the fab cycle time in days versus the utilization of the fab (%). The utilization of the fab is defined as the WIP divided by the total capacity of the fab. 

We have found this method useful in understanding the fundamental principles of cycle time. The operating curve helps to explain how factory physics impact fab KPIs such as cycle time and fab utilization by showing the changes in the operating points: 

• The horizontal line, representing the summation of raw process times (known as theoretical cycle time), envisions a scenario with zero queuing time in the fab. This illustrates the impact of queuing time on cycle time as we increase WIP, moving right on the x-axis. Accumulation of queuing time becomes inevitable with the introduction of more WIP in the fab.
  ‍
• The ideal operating curve represents the operation of a wafer-fab assuming that there is zero waste. This curve defines the minimum achievable cycle time for each fab load and the difference between this curve and theoretical cycle time is because of real life variability in the fab that cannot be eliminated completely, e.g. unplanned maintenances, inconsistent tool processing times.
  ‍
• The cycle time tends to go to infinity, when you move towards 100% utilization of the fab.
  ‍
• The actual operating curve, cycle time versus factory utilization, represents the current fab’s operation considering all the inefficiencies such as excessive inventory, variability in operations, idle times, poor batching and rework.
  ‍
• Both curves assume average or constant values of the operational parameters of the fab for example a fixed number of tools installed, an average availability of each tool and labor, and a constant product mix.
  ‍
• The actual operating curve describes the impact on cycle time if we load the fab with more WIP as shown in Figure 4. The fab management could use this information to make a decision about the trade-off between cycle time and factory utilization. Higher fab utilization is associated with a higher throughput (i.e. number of wafers per unit of time).

In Figure 3, you can see that the current fab cycle time is 40 days when the factory utilization is at 60%. Theoretically, we could reduce the cycle time to 22 days. The difference between these two points is due to the inefficiencies that contribute to the factory cycle time as explained in the introduction of this section. In Part 2 of this blog, we will explore the various types of inefficiencies and examine how innovation can shift the operating curve to achieve lower cycle times while maintaining the same fab utilization.

Summary 

In summary, cycle time is not merely a production metric but the very pulse of the semiconductor manufacturing and supply chain. It governs revenues, shapes market responsiveness, and is pivotal in driving innovation. By understanding its nuances, semiconductor companies can not only optimize their operations but also gain a competitive edge. And while we've scratched the surface on its significance, the question remains: how can we further reduce and refine it? In part 2 of the C for Cycle Time blog, we will discover innovative techniques that promise to revolutionize cycle time management in wafer fabs.

‍

Author: Dennis Xenos, CTO and Cofounder, Flexciton

References

• [1] James P. Ignizio, 2009 ,Optimizing Factory Performance: Cost-Effective Ways to Achieve Significant and Sustainable Improvement 1st Edition, McGraw-Hill, ISBN 978-0-07-163285-0
• [2] Semiconductor Industry Association, 2021, Blog, URL
• [3] Deenen, P.C., Middelhuis, J., Akcay, A. et al., 2023, Data-driven aggregate modeling of a semiconductor wafer fab to predict WIP levels and cycle time distributions. Flex Serv Manuf J. https://doi.org/10.1007/s10696-023-09501-1
• [4] Ondrej Burkacky, Marc de Jong, and Julia Dragon, 2022, Strategies to lead in the semiconductor world, McKinsey Article, URL. 

‍

Welcome to a nuanced exploration of pivotal considerations surrounding cloud adoption in the context of wafer fabrication. For those reading sceptically, uncertain about the merits of cloud integration, or perhaps prompted by concerns about lagging behind competitors—this blog endeavours to shed light on key areas of relevance.

Introduction

For those reading this blog, the chances are you (or perhaps your boss) remain unconvinced about the merits of cloud adoption, yet are open to participating in the ongoing debate. Alternatively, there might be a concern of falling behind industry peers, perhaps heightened by recent security incidents such as the hacking of X-Fab. By the end of this short article, you will have gained valuable insights into the significant areas of cloud security, with the anticipation that such information will contribute to a more informed decision-making process.

Firstly, this is about using a cloud service, not running your own systems in the cloud. There are good arguments for that too, but that’s not what this article is about. So, the areas deemed worthy of exploration within this context include:

• Security - It is a paramount concern that influences the reluctance of fabs to embrace cloud services, often accompanied by apprehension about entrusting sensitive data to cloud platforms.
• Type of service - If you’re running it on-premise, you’re probably managing it. Would you prefer to be buying software or a service?
• Cost - One of the supposed benefits of the cloud is less overall running cost. How much truth is in that? 
• Reliability and criticality - What about backups and disaster recovery and failover and downtime and maintenance and…?
  
  ‍

Recognising the complexity of these topics, we aim to take a segmented approach, with this blog dedicating its focus to the critical factor of security. Subsequent entries promise a comprehensive discussion on the remaining aspects. 

Security

We’re going to start with a simple one. Is your fab in any way connected to the internet? If you’re genuinely air-gapped, then it's reasonable to assume you already have a high level of security. But, if you’re not actually air-gapped, then you could actually improve your security by using a cloud service rather than running that service on-prem. Not instantly obvious perhaps, but let us explain. 

My fab is connected to the internet already

The most compelling argument that exists for this is a simple one. Microsoft, AWS, IBM and Google all run respectable professional public clouds. If the service we’re talking about connecting to runs on any one of them, it’s fair to say they have similar approaches to cybersecurity.

Microsoft alone employs 3500 cybersecurity professionals to maintain the security of Azure and together they spend a lot on cybersecurity improvements. That’s an awful lot more person-hours on security than most are going to be able to apply from their team. Every single one of them is contributing to the security of a system running in their cloud.

“Aha!”, you say, “that tells me that the underlying public cloud infrastructure that the service is running on is probably as secure as anything connected to the world could be, but that doesn't mean that the service running on it is, right?” And yes, that’s a fair concern. As one of those service providers, we can confirm that we do not employ 3500 cybersecurity professionals. But because we run our service on Azure, we don’t need to. More than half our fight is already done for us and the remainder is a lot easier. For example:

• Staying on top of security patches: Not something we need to worry about. Azure does almost all of it for us. The only thing it doesn’t do is update our own code dependencies, so we have Snyk do that for us.
• Anti-virus: Don’t want viruses in your systems? Tick. That’s an easy built-in feature.
• Secure data: All storage is encrypted without us having to do anything special and we can do the same with data in transit.
• Secret management: There’s a whole infrastructure designed by those 3500 cybersecurity professionals that we can use right off the shelf.
• Access management: Active Directory–a directory service built by Microsoft–is built into everything so we just have to define what the roles should be and who has access. There’s not many roles, and only some very trusted people.
• Security auditing: Every action in every piece of infrastructure is logged by default and we can feed that all in very easily to our security monitoring systems to identify any suspicious activity.
  
  ‍

In discussing the ease of these security measures, perhaps we’ve been slightly frivolous. However, despite the casual tone, the implementation of security measures when using cloud technologies is notably simpler when compared with organisations that manage their own hardware. 

My fab is air-gapped right now

On the other hand, maybe you’re a fab that is actually air-gapped. You’ve got a solid on-site security team and excellent anti-social-engineering measures. Why introduce any risk? Fair question. We’d argue that this is going to become an increasingly challenging problem for you and maybe now’s the time to get ahead of the problem. Tools on your shop floor are already getting more modern, with virtualised metrology and off-site telemetry feeds for predicting failure rates using machine learning. Some of these systems just can’t be run on site and you’ll increasingly have to do without the more advanced aspects of your tooling to maintain your air gap. Over time this will take its toll, and your competitors will begin to pull away.

At this point it’s worth mentioning that SEMI has put together standards in the cybersecurity space. These address risks like bringing tools into your network with embedded software on them as well as defining how to set up your fab network to secure it, while still enabling external communication. We’d suggest that you should treat a cloud service no differently. It is entirely possible to use a managed service, in the cloud, connected to your fab, while still relying on purely outbound connectivity from your fab, leaving you entirely in control of what data is provided to the service and what you do with any data made available by that service in return.

In summary

If you’re already “internet-enabled” in your fab, then we’d argue that using a reputable public cloud service is actually more secure than running that same service on-prem.

If you’re completely offline, we’re not going to argue that using a cloud service is more secure than not connecting to the internet. What I am arguing though, is that at some point you’re going to have to anyway, so you’re better off getting on top of this now rather than waiting until you’re forced into it by the market. 

‍

Author: Ray Cooke, VP of Engineering at Flexciton

‍

For many years, my career has been deeply rooted in the ever-changing world of manufacturing–an industry where progress relies on innovation. Throughout my professional journey, I have been immersed in this dynamic sector, focusing on creating bespoke software solutions for manufacturing and logistics, all the while seamlessly integrating third-party solutions into established workflows. My experience has afforded me the opportunity to first-hand witness the profound changes that digitalisation and automation have brought to the manufacturing landscape. As technology and manufacturing processes have become more closely intertwined, the operational dynamics of production have been reshaped.

Like any successful partnership, the marriage of manufacturing and technology requires a strong foundation built on trust, mutual understanding, respect, and a shared ambition to support each other's growth and empowerment. However, these transformative shifts have brought along their fair share of challenges and concerns that continue to echo around the manufacturing world. 

Embracing servant software in the manufacturing landscape

A few years ago, I collaborated with a couple of value stream managers as we scoured the market for various digital products, seeking the optimal solution to integrate with our in-house developed material requirements planning (MRP) system. 

One significant concern was the fear of adopting software that was too intrusive. In an industry where precision and control are paramount, the idea of software delving too deeply into our operations was disconcerting. Even worse was the fear of getting locked into specific technologies. Having deeply integrated software within our operations poses a risk due to them being so costly to replace, which potentially limits our capacity to adapt and evolve in tandem with the industry. We wanted automation and the ability to forecast the incoming work. Our aim was to prevent defects and misjudgments, all the while ensuring that we retained control over our manufacturing processes. And importantly, we were adamant about not compromising our quality standards.

The reality is that the market for manufacturing-oriented software is littered with solutions that are cumbersome, inflexible, and expensive. When I joined Flexciton as a Senior Product Manager, I was pleasantly surprised to discover a refreshing departure from the norm in Flexciton's product philosophy. 

It evokes the concept of “servant software”. Similar to the idea of servant leadership–where a leader prioritises the well-being, growth and empowerment of team members–servant software aims to streamline processes, simplify tasks, and provide solutions that cater to the users' requirements and preferences. 

A servant software encompasses, as a foundational principle, the advantage of being as flexible and adaptable as a meticulously tailored suit. This quote summarises the concept: 

Upgrade your user, not your product. Don’t build better cameras — build better photographers.

— Kathy Sierra

In the challenge of digitising semiconductor wafer fabs, Flexciton aspires to play a pivotal role in cultivating highly skilled operators and managers—individuals who are empowered by our technology rather than being replaced by it.‍

Automation for enrichment, not alienation

Picture Josh, a Senior Fab Operator in the diffusion area, who has been working for five years in a manually operated wafer fab. Half of his workday is consumed by the arduous task of sifting through a colossal spreadsheet that meticulously logs all the lots in progress, each with its own unique characteristics. He sits at his desk, constantly toggling between this spreadsheet and another monitor displaying the real-time status of the tools.

Jotting down notes on a piece of paper, Josh ventures into the tangible world of the fab. There, he confronts the actual events unfolding. He asks himself, "Is this an actuality? Are these lots genuinely ready for processing? Can I really preload this tool?" Realisation strikes: "No, they are still in transit, and I cannot proceed with this batch," or "I can’t preload this tool yet; a few minutes are still left." Josh retreats to his desk to recalibrate his plans once more.

When operators are liberated from repetitive and inefficient tasks, they can harness their cognitive abilities to identify improvement opportunities, propose innovative solutions, and implement process enhancements directing their efforts towards value-added activities that demand uniquely human qualities. This empowerment not only enhances job satisfaction but also drives a culture of ownership and accountability.

Embracing Lean Management principles

Servant software aligns seamlessly with the principles of lean management, a philosophy that champions efficiency through the elimination of waste and continuous improvement. Lean management is not just about operational optimization, it emphasises a shift in mindset, encouraging all levels of an organisation to work cohesively towards shared objectives. By integrating servant software within this framework, manufacturers can elevate their workforce's role away from simply executing tasks and towards contributing to the bigger picture.

Operators typically concentrate their efforts within their designated areas of responsibility, striving to optimize operations by carefully managing various tasks. They work diligently to maintain a delicate balance among tools, ensuring workloads are efficiently allocated, changeovers are optimized, and maintenance and process control activities are accommodated for. Even within a confined production area, this manual juggling of numerous constraints and variables presents a considerable challenge, a topic we explored further in our article on autonomous scheduling. 

A new way to schedule the fab is the key. But what’s in it for the operators? What is the impact on their daily work? Our software aims to provide operators with a tool that leads them to take the right action at precisely the right moment. It ensures that tasks are executed with impeccable timing, neither prematurely nor delayed, considering not only the current status of the WIP (work in progress) and the tools they are responsible for, but also the potential effect of their actions on the following production stages.
‍
This goes beyond optimizing individual areas; instead, it is designed to harmonise the entire manufacturing process. By avoiding over-optimization of one area, we prevent potential bottlenecks or resource shortages elsewhere in the workflow, resulting in a balanced, easily monitored, and controllable production process.

The fab in your pocket!

Our operators' tools are integral to the Flexciton application ecosystem, where every component is integrated and consistent. From analytics and scheduling to automated tuning, and extending to the practical, hands-on actions of our operators—such as loading or unloading tools or conducting Statistical Process Control (SPC) tasks—our system comprehensively covers all aspects. Therefore, Josh can simply glance at his portable device to discern the next best action to perform or be notified when something urgently requires his attention.

Our primary goal is to provide operators with the essential information they need, without overwhelming them. This information is easily accessible on portable devices, ensuring its effectiveness from the very first day an operator steps into the fab.
‍
Operators—now armed with useful insights and empowered by automation—can expand their contributions beyond their individual roles, engaging in more value-adding tasks. The result is a collaborative ecosystem where every individual becomes a key player in achieving fab-wide targets and goals.

‍

‍

A customer-centric product philosophy

In delivering software solutions for the semiconductor industry, our mission revolves around achieving an optimal balance, thereby cultivating a modern, flexible, and customer-centric product philosophy. Our platform, while robust, maintains a deep respect for operational boundaries, ensuring that our customers are not confined to rigid models.

Instead, it functions as a dynamic tool that enriches adaptability and innovation, and grants users complete control over their manufacturing processes. By adhering to these core principles and relentlessly pursuing software that empowers without overwhelming, we unlock the full potential of a harmonious synergy between technology and manufacturing, propelling progress forward without concessions.

‍

Author: Valentina Vivian, Senior Product Manager at Flexciton

Welcome back to the Flexciton Tech Glossary Series: A Deep Dive into Semiconductor Technology and Innovation. Our second entry of the series is all about Batching. Let's get started!

A source of variability

Let's begin with the basics: what exactly is a batch? In wafer fabrication, a wafer batch is a group of wafers that are processed (or transported) together. Efficiently forming batches is a common challenge in fabs. While both logistics and processing both wrestle with this issue, our article will focus on batching for processing, which can be either simultaneous or sequential.

Simultaneous batching is when wafers are processed at the same time on the same machine. It is very much inherent to the entire industry, as most of the machines are designed for handling lots of 25 wafers. There are also process types – such as thermal processing (e.g. diffusion, oxidation & annealing), certain deposition processes, and wet processes (e.g. cleaning) – that benefit from running multiple lots in parallel. All of these processes get higher uniformity and machine efficiency from simultaneous batching.

On the other hand, sequential batching refers to the practice of grouping lots or wafers for processing in a specific order to minimise setup changes on a machine. This method aims to maximise Overall Equipment Effectiveness (OEE) by reducing the frequency of setup adjustments needed when transitioning between different production runs. Examples in wafer fabrication include implant, photolithography (photo), and etch. 

Essentially, the entire process flow in wafer manufacturing has to deal with batching processes. To give a rough idea: a typical complementary metal-oxide semiconductor (CMOS) architecture in the front-end of the line involves batching in up to 70% of its value added steps. In a recent poll launched by FabTime on what the top cycle time contributors are, the community placed batching at number 5^[1], behind tool downs, tool utilisation, holds, and one-of-a-kind tools. Batching creates lot departures in bursts, and hence it inherently causes variability in arrivals downstream. Factory Physics states that:

“In a line where releases are independent of completions, variability early in a routing increases cycle time more than equivalent variability later in the routing.” ^[2]

Successfully controlling this source of variability will inevitably result in smoother running down the line. However, trying to reduce variability in arrival rates downstream can lead to smaller batch sizes or shorter campaign lengths, affecting the effectiveness of the batching machines themselves.

The many complexities of batching

In wafer fabs, and even more so in those with high product mix, batching is particularly complicated. As described in Factory Physics:

"In simultaneous batching, the basic trade-off is between effective capacity utilisation, for which we want large batches, and minimal wait to batch time, for which we want small batches.” ^[2]

For sequential batching, changing over to a different setup of the machine will cause the new arriving lots to wait until the required setup is available again.

In both cases, we’re talking about a decision to wait or not to wait. The problem can easily be expressed mathematically if we’re dealing with single product manufacturing and a low number of machines to schedule. However, as one can imagine, the higher the product mix, the higher the possible setups and machines. Then the problem complexity increases, and the size of the solution space explodes. That’s not all, there are other factors that might come into play and complicate things even more. Four different examples are:

• Timelinks or queue time constraints: a maximum time in between processing steps
• High-priority lots: those that need to move faster through the line for any reason
• Downstream capacity constraints: machines that should not get starved at any cost
• Pattern matching: when the sequence of batching processes need to match a predefined pattern, such as AABBB

Strategies to deal with batching

Historically, the industry has used policies for batching; common rules of thumb that could essentially be split up into ‘greedy’ or ‘full batch’ policies^[3]. Full batch policies require lots to wait until a full batch is available. They tend to favour effective capacity utilisation and cost factors, while they negatively impact cycle time and variability. Greedy policies don’t wait for full batches and favour cycle time. They assume that when utilisation levels are high, there will be enough WIP to make full batches anyway. For sequential batching on machines with setups, common rules include minimum and maximum campaign length, which have their own counterpart configurations for greedy vs full batching.^[3] 

The batching formation required in sequential or simultaneous batching involves far more complex decisions than that of loading a single lot into a tool, as it necessitates determining which lots can be grouped together. Compatibility between lots must be considered, and practitioners must also optimize the timing for existing lots on the rack to await new arrivals, all with the goal of maximising batch size. ^[4]

Industrial engineers face the challenge of deciding the best strategy to use for loading batch tools, such as those in the diffusion area. In an article by FabTime ^[4], ^[5] the impact of the greedy vs full or near full batch policy is compared. The greedy heuristic reduces queuing time and variability but may not be cost-effective. Full batching is cost-effective but can be problematic when operational parameters change. For instance, if a tool's load decreases (becomes less of a bottleneck), a full batch policy may increase cycle time and overall fab variability. On the other hand, a greedy approach might cause delays for individual lots arriving just after a batch is loaded, especially critical or hot lots with narrow timelink windows. Adapting these rules to changing fab conditions is essential.

In reality, these two policies are extreme settings in a spectrum of possible trade-offs between cost and cycle time (and sometimes quality). To address the limitations of both the greedy and full batch policies, a middle-ground approach exists. It involves establishing minimum batch size rules and waiting for a set duration, X minutes, until a minimum of Y lots are ready for batching. This solution usually lacks robustness because the X and Y values depend on various operational parameters, different recipes, product mix, and WIP level. As this rule-based approach incorporates more parameters, it demands greater manual adjustments when fab/tool settings change, inevitably leading to suboptimal tool performance.

In all of the above solutions, timelink constraints are not taken into consideration. To address this, Sebastian Knopp^[6] recently developed an advanced heuristic based on disjunctive graph representation. The model's primary aim was to diminish the problem size while incorporating timelink constraints. The approach successfully tackled real-life industrial cases but of an unknown problem size.

Over the years, the wafer manufacturing industry has come up with various methodologies to help deal with the situation above, but they give no guarantee that the eventual policy is anywhere near optimal and their rules tend to stay as-is without adjusting to new situations. At times, this rigidity has been addressed using simulation software, enabling factories to experiment with various batching policy configurations. However, this approach proved to be resource-intensive and repetitive, with no guarantee of achieving optimal results.

How optimization can help master the batching problem

Optimization is the key to avoiding the inherent rigidity and unresponsiveness of heuristic approaches, helping to effectively address the batching problem. An optimization-based solution takes into account all batching constraints, including timelinks, and determines the ideal balance between batching cost and cycle time, simultaneously optimizing both objectives.

It can decide how long to wait for the next lots, considering the accumulating queuing time of the current lots and the predicted time for new lots to arrive. No predetermined rules are in place; instead, the mathematical formulation encompasses all possible solutions. With a user-defined objective function featuring customised weights, an optimization solver autonomously identifies the optimal trade-off, eliminating the need for manual intervention.

The challenge with traditional optimization-based solutions is the computational time when the size and complexity of the problem increase. In an article by Mason et al.^[7], an optimization-based solution is compared to heuristics. While optimization outperforms heuristics in smaller-scale problems, its performance diminishes as problem size increases. Notably, these examples did not account for timelink constraints.

This tells us that the best practice is to try to break down the overall problem into smaller problems and use optimization to maximise the benefit. At Flexciton, advanced decomposition techniques are used to break down the problem to find a good trade-off between reduced optimality from the original problem and dealing with NP-hard complexity.^[8]

Many practitioners aspire to attain optimal solutions for large-scale problems through traditional optimization techniques. However, our focus lies in achieving comprehensive solutions that blend heuristics, mathematical optimization, like mixed-integer linear programming (MILP), and data analytics. This innovative hybrid approach can vastly outperform existing scheduling methods reliant on basic heuristics and rule-based approaches.

Going deeper into the solution space

In a batching context, the solution space represents the numerous ways to create batches with given WIP. Even in a small wafer fab with a basic batching toolset, this space is immense, making it impossible for a human to find the best solution in a multi-product environment. Batching policies throughout history have been like different paths for exploring this space, helping us navigate complex batching mathematics. Just as the Hubble space telescope aided space exploration in the 20th century, cloud computing and artificial intelligence now provide unprecedented capabilities for exploring the mathematical world of solution space, revealing possibilities beyond imagination.

With the advent of these cutting-edge technologies, it is now a matter of finding a solution that satisfies the diverse needs of a fab, including cost, lead time, delivery, quality, flexibility, safety, and sustainability. These objectives often conflict, and ultimately, finding the optimal trade-off is a business decision, but the rise of cloud and AI will enable engineers to pinpoint a batching policy that is closest to the desired optimal trade-off point. Mathematical optimization is an example of a technique that historically had hit its computational limitations and, therefore, its practical usefulness in wafer manufacturing. However, mathematicians knew there was a whole world to explore, just like astronomers always knew there were exciting things beyond our galaxy. Now, with mathematicians having their own big telescope, the wafer manufacturers are ready to set their new frontiers.

‍

Authors
Ben Van Damme, Industrial Engineer and Business Consultant, Flexciton
Dennis Xenos, CTO and Cofounder, Flexciton

References

[1] FabTime Newsletter: Issue 24.03

[2] Wallace J. Hopp, Mark L. Spearman, Factory Physics: Third Edition. Waveland Press, 2011

[3] Lars Mönch,  John W. Fowler,  Scott J. Mason, 2013, Production Planning and Control for Semiconductor Wafer Fabrication Facilities, Modeling, Analysis, and Systems, Volume 52, Operations Research/Computer Science Interfaces Series 

[4] FabTime Newsletter: FabTime Cycle Time Tip of the Month #4: Use a Greedy Policy when Loading Batch Tools

[5] FabTime Newsletter: Issue 9.03 

[6] Sebastian Knopp, 2016, Complex Job-Shop Scheduling with Batching in Semiconductor Manufacturing, PhD thesis, l’École des Mines de Saint-Étienne 

[7] S. J. Mason , J. W. Fowler , W. M. Carlyle & D. C. Montgomery, 2007, Heuristics for minimizing total weighted tardiness in complex job shops, International Journal of Production Research, Vol. 43, No. 10, 15 May 2005, 1943–1963   

[8] S. Elaoud, R. Williamson, B. E. Sanli and D. Xenos, Multi-Objective Parallel Batch Scheduling In Wafer Fabs With Job Timelink Constraints, 2021 Winter Simulation Conference (WSC), 2021, pp. 1-11

‍

To cloud, or not cloud, that is the question.

Some might consider the opening statement a tad flippant in borrowing Hamlet's famous soliloquy. Yet, the internal struggle our hero feels agonising over life and death holds a certain likeness to the challenges faced by Fab Managers today. Businesses live and die by their decisions to either embrace or disregard new innovations to gain a competitive edge and nowhere is this truer than in the rough and tumble world of semiconductor manufacturing; Fairchild, National Semiconductor and S3 are just a few of those who did not last. ^[1][2][3]‍

Semiconductor manufacturing has had a long history of innovating, tweaking, and tinkering,^[4] so it’s somewhat surprising that the sentiment towards cloud uptake has been weaker in the semiconductor industry compared to the wider market^[5]. This article aims to explore some of the potential benefits of cloud adoption to better equip Fab Managers with the motivation to take another look at the cloud question.

Recap: What are the different types of Cloud?

Cloud computing encompasses public, private, and hybrid models. The public cloud (think Azure, AWS, Google Cloud and so on) offers rental of computational services over the internet, while the private cloud replicates cloud functionality on-premises. However, private clouds require a significant upfront investment, ongoing maintenance costs and a skilled in-house IT team to manage and maintain the infrastructure, making it a less appealing option for smaller firms. Hybrid cloud blends on-site and cloud resources for flexible workloads, segregating the most sensitive workloads to on-premise environments for the greatest control; however, control does not necessarily mean security, which will be discussed in a later article! 

Understanding the benefits of cloud

1.      The Latest Tech

Embracing the latest cloud technology offers wafer fab facilities, not just organisations, a direct path to heightened capabilities in their manufacturing processes through the use of digital and smart manufacturing technologies. By harnessing advanced computational powers such as real-time analytics; optimization^[6]; and machine learning defects detection^[7], fabs can maximise all their fundamental KPIs, ultimately leading to better business outcomes. McKinsey estimates that, compared to other semiconductor activities, manufacturing has the most to gain from the AI revolution (Fig. 1), and a key technology enabling this is will be the vast computational power of the cloud.^[8]

‍

Case Study: The Latest Tech Driving Improvements in Fab KPIs

Seagate achieved a 9% increase in moves by utilising Flexciton’s cloud native platform and cutting-edge autonomous scheduling.

2. Redundancy, Scaling, Recovery and Updates

It is true that some of these technologies can be provided on-premises; however, cloud computing, in general, reduces downtime through redundancy, automated scaling, and disaster recovery mechanisms, ensuring seamless operation even during hardware failures or unexpected traffic spikes. Some estimates suggest that downtime can cost firms an eye-watering $1 million to $5 million per hour, depending on their size and sector. ^[9] By leveraging the cloud, the cost of operating disaster recovery services has demonstrated potential cost savings of up to 85% when comparing public to private options. ^[10] It is easy to speculate that for wafer fab critical infrastructure, the cost of downtime could be significantly higher.

Furthermore, the number of wafers processed within a fab can cause computational traffic spikes during busy periods for some applications. On-premises deployments would need to account for this, even if the resource is not in use all the time, which can add to inefficiencies, while public cloud can elastically scale down, meaning you only pay for what you use. 

Lastly, on-premises systems without the ability to monitor and update remotely are often many versions behind, prioritising perceived stability but research has shown increasing the rate of software iteration increases stability and resilience rather than weakening it. ^[11] Without the convenience of remote updates, legacy systems can become entrenched, with employees on the shop floor being hesitant to embrace change due to the fear of disrupting critical infrastructure and the expenses associated with upgrading IT infrastructure. This sets in motion a self-reinforcing cycle where the expenses and associated risks of transitioning increase over time, ultimately resulting in significant productivity losses as users continue to rely on outdated technology from decades past.

3. Specialisation and Comparative Advantage

Stepping back from the fab and taking a holistic view of the semiconductor manufacturing organisation reveals compelling economic arguments, both on macro and micro scales, for embracing cloud.

Allowing cloud providers to specialise in cloud computing while wafer fab manufacturers focus solely on wafer fabrication benefits the latter by freeing them from the complexities of managing IT infrastructure. ^[12] This collaboration allows wafer fab manufacturers to allocate their resources towards core competencies, leading to increased operational efficiency and superior wafer production.

Simply put, fabs do not build the complex tools they need to make their products, such as photolithography equipment; they purchase and utilise them in ways others can’t to produce market leading products. Why should utilising the tools of the cloud be any different?

On a macro level, the argument of specialisation also applies through comparative advantage.^[13] Different continents and countries have comparative advantages in certain fields, Asia has long been a world leader in all kinds of manufacturing due to its vast populations.^[14] The United States, on the other hand, has a tertiary education system which is the envy of the world; institutions like Stanford and MIT are household names across the globe, and this has provided the high technical skills needed to be the home of the technology start up. Utilisation of cloud technology and other distributed systems allows firms to take the best of what both regions have to offer, high tech manufacturing facilities from Singapore to Taiwan with the latest technology from Silicon Valley or perhaps London. Through the cloud, Fab Managers and organisations can leverage a single advanced technology across multiple fabs within complex supply chains. This eliminates the need for costly and experienced teams to travel across the globe or manage multiple teams in various locations with varying skill sets, all while locating facilities and offices where the best talent is.

In brief, semiconductor firms' fate could rest on one pivotal decision: adoption of cloud. This choice carries the promise of leveraging cutting-edge technology, fortifying resilience, and reaping a multitude of advantages. Notably, by transitioning to cloud-native solutions, Fab Managers can usher their organisations into an era of unparalleled competitiveness, all while enjoying a range of substantial benefits. Among these benefits, for example, is cloud-native architecture like Flexciton’s, promising lower cost of ownership and zero-touch maintenance for fabs. We will delve deeper into the crucial aspect of security in one of our upcoming blogs, providing a comprehensive understanding of how cloud-native solutions are actually key to safeguarding sensitive data and intellectual property, rather than compromising it. In this era of constant innovation, embracing the cloud is more than just an option; it’s becoming a strategic imperative.

Author: Laurence Bigos, Product Manager at Flexciton

‍

References

[1] Investor relations - Texas Instruments completes acquisition of National Semiconductor - Texas Instruments

[2] ON Semiconductor Successfully Completes Acquisition of Fairchild Semiconductor for $2.4 Billion in Cash

[3] S3 Graphics: Gone But Not Forgotten | TechSpot

[4] Miller, C. (2022). Chip War: The Fight for the World's Most Critical Technology. Scribner.

[5] Flexciton | Blog & News | Is Fear Holding Back The Chip Industry’s Future In The Cloud?

[6] Flexciton | Resources | Seagate Case Study 2.0

[7] Lynceus: Inline, Real-time, AI Based Process Control Monitoring That Can Reduce Inspection & Metrology Capex (semianalysis.com)

[8] Applying artificial intelligence at scale in semiconductor manufacturing | McKinsey

[9] Know Key Disaster Recovery Statistics And Save Your Business (invenioit.com)

[10] Wood.pdf (usenix.org)

[11] Forsgren, N., Humble, J., & Kim, G. (2018). Accelerate: The Science of Lean Software and DevOps. IT Revolution Press.

[12] Specialization Definition (investopedia.com)

[13] What Is Comparative Advantage? (investopedia.com)

[14] Why China Is "The World's Factory" (investopedia.com)

‍

At Flexciton, we often talk about how autonomous scheduling allows wafer fabs to surpass the need for maintaining many rules to enable the behaviours they want at different toolsets. I would like to offer an analogy to show how significant the difference is.

Navigating the City

Imagine you are a passenger in a taxi. Your driver is a local; they know every road like the back of their hand and know the best routes to avoid likely problems. They can be flexible and effective, but have to spend a long time thinking about how to get to your destination. They also can’t know about the traffic on each potential route, and for new destinations they may require some trial and error before they find a good way of getting there. Worst of all, though they might have accumulated some great stories from their years of driving, it’s only thanks to those many years that they can navigate with any level of mastery.

Now imagine you have a very basic robotic driver; this driver is so mechanical that it has a hard-coded rule for every single road and junction: “If I’m at junction 20, I wait exactly thirty seconds and then I turn left.” This rule has come from an engineer performing a time study based on traffic levels six months ago. The driver has no knowledge of local events happening (for example, if it turns out that there is no oncoming traffic right now), and doesn’t even change its decisions when you need it to navigate to a new destination!

Meanwhile, when local conditions change at all (gaps in oncoming traffic at junction 20 are now every twenty seconds on average!) an engineer needs to manually change that parameter in the robot’s logic. And if the overall conditions change everywhere, or a new destination is desired, every rule needs to be retuned.

‍

Finally, imagine a truly autonomous taxi. This taxi has a navigation system that knows where the traffic is, assesses the speed of every potential route, reacts to changes in conditions, and can get you to exactly where you want to go. In fact, all you have to do is tell it the destination; then you can sit back and relax, knowing it will get you there in the shortest possible time.

Navigating the Fab

While many wafer fabs have moved away from relying purely on tribal knowledge of manufacturing specialists on the fab floor, the scheduling problem in semiconductor factories is so difficult that, until recently, the hard-coded robotic taxi driver was the state-of-the-art. These solutions ask industrial engineers to manually tune thousands of rules to achieve intelligent behaviour, and they must be continuously re-tuned as fab conditions change.

A common scheduling challenge is deciding when to allow wafers into a timelink (or queue time loop) at diffusion. A timelink is the maximum amount of time that can elapse between two or more consecutive process steps, and some schedulers will simply limit the number of lots allowed within the timelinked steps at any one time. Others will just use a priority weight given to all timelinked lots, so that they are more likely to move through the loop without violating their time limit. Both of these rules are manually tuned and can’t react to the conditions of that particular moment in time, leading either to rework or scrap, or unnecessarily high cycle times.

Another typical example from a commonly-used heuristic scheduler is the application of minimum batch size rules at diffusion areas. A typical rule might be “wait for a minimum batch size of x, unless y minutes have elapsed, in which case dispatch whatever is at the rack.” Many fabs will set up this rule for every furnace-operation combination, which could mean ~3,000 manually tuned parameters just for one rule at one toolset.

Meanwhile, when micro conditions change, for example daily wip level fluctuations, these tuned parameters cannot react. And worse, when macro conditions such as overall market demand change, it makes it very hard for the whole fab to pivot quickly, because every rule needs re-tuning manually. Despite the theory that these rules can be set once every few months, in practice most fabs end up re-tuning these rules continuously, even daily, in order to maintain reasonable performance - accepting the predictable impact on industrial engineering resources that has!

Optimized scheduling, however, does away with these rules entirely and directly calculates the optimal schedule to improve your chosen objectives. In the timelink example, it doesn’t need to rely on guessing how many lots can be allowed into the loop - it just calculates the optimal schedule for the multiple steps involved, ensuring no timelink violations will occur.

But still, how do you get the scheduler to do what you want?

A New Paradigm for Tuning

If you have read any of our previous articles, you may be aware that optimization-based scheduling uses objectives such as “minimise queue time” and “maximise batch size” to calculate the optimal schedule. In fact, on most of our toolsets we only use ~2-3 objective weights, and by setting these you can achieve the balance and results you want.

Even this, however, is not truly autonomous.

We’ve been working to bring forward a new paradigm: letting you choose the fab-level outcome you want directly - like setting the destination for the taxi. If you know you want to prioritise achieving higher throughput, you can just specify that and Flexciton’s autonomous scheduler will automatically figure out what the optimization objective needs to be to achieve it.

What does this mean? It means you directly control the fab outcome you want to achieve, rather than guessing what toolset-level behaviours will produce the fab-level KPIs you want.

Orders of Magnitude

So when we speak about autonomous scheduling, we are referring to this new paradigm where you can choose the outcome you want, and Flexciton automatically does the rest. Instead of ~3,000 manually tuned parameters for just one of many rules at one toolset, just pick your desired KPI trade-off, and we automatically set the handful of objective weights that drive the optimization engine.

The result is not needing industrial engineering resource dedicated to tuning – instead, this valuable resource can be redeployed to work on higher value tasks. Moreover, it can enable consistent high performance across changes in fab conditions; and it becomes easier to pivot the entire fab’s direction when market conditions change.

This is how Flexciton’s scheduler is powerful enough to let you set the destination, and go.

‍

Author: Sebastian Steele, Product Manager at Flexciton

We are excited to introduce the Flexciton Tech Glossary Blog Series: A Deep Dive into Semiconductor Technology and Innovation.

In an ever-evolving semiconductor industry, understanding the nuances of new technologies and the transformative potential of artificial intelligence and optimization is paramount. The Flexciton Tech Glossary Blog Series is designed to shed light on specific technologies and innovations, offering insights into how these advancements can revolutionise semiconductor manufacturing operations.

Each article in this series will delve into a distinct theme, aiming to equip any practitioner in the industry from industrial engineers and manufacturing experts up to VP level professionals with the knowledge to integrate these innovations into their daily operations.

Beyond our in-house expertise, we’re excited to collaborate with industry experts, inviting them to contribute and enrich our series with their specialised knowledge and experience. Join us on this enlightening journey as we explore the frontiers of the semiconductor industry from A-Z.

AI will transform the semiconductor industry

Artificial Intelligence (AI) has become a transformative force in various sectors, driving a global wave of innovation and automation. Seemingly overnight, systems like ChatGPT that harness the primary human interface – natural language – have revolutionised how we interact with technology. In a similar vein, generative art technologies have reinvented our relationship with creativity, making it more accessible than ever before. These remarkable systems have acquired their capabilities through learning, fueled by training on vast amounts of data. This ongoing revolution prompts the question: what is the next frontier to be conquered?

Beyond the novel consumer applications leading the charge, the implications of AI in specialised fields, such as semiconductor manufacturing, are equally profound. Estimates place the earnings already achieved by AI across the semiconductor value chain at over $5 billion. The range of applications is immense and spans activities at all levels. From informing capital allocation, to demand forecasting, fab layout planning, and right down to chip design, AI can enable automation and increase efficiency. Semiconductor manufacturing, in particular, has been identified as the function presenting the most attractive opportunities, where the potential savings have been calculated to be over $10 billion in just the next few years [1].

Impact at all levels 

The semiconductor industry is facing several challenges where AI can make a significant impact. These span all the industry’s key activities: long-term capacity planning, research & design, sales, procurement and, of course, manufacturing. Some use cases that are increasingly gaining traction are:

• Supply Chain Optimization: Predictive analytics can forecast demand, optimize inventory levels, and enhance the overall efficiency of the supply chain [2]. 
• Automated Material Handling Systems (AMHS): Utilising AI-driven cognitive robotics within AMHS automates material transportation throughout the plant [3], optimizing production planning considering AMHS [4].
• Predictive Maintenance: AI can predict when equipment is likely to fail or require maintenance, reducing downtime and increasing overall equipment efficiency [5].
• Defect Detection: Advanced image recognition algorithms can identify defects in wafers at an early stage, ensuring higher yields and reducing wastage [6].
• Virtual metrology: AI can be deployed to estimate a product’s quality directly from production process data. This enables real-time quality monitoring without additional measuring steps [7]. 
• Process control: AI can analyse vast amounts of data to optimize the manufacturing process, ensuring the best conditions for each step and improving the overall quality of the chips (e.g. tool matching) [8].
  ‍
  ‍

In this article, we focus on AI’s potential to automate scheduling within a semiconductor wafer fab and improve key metrics: increase the throughput of manufacturing lines, reduce cycle times and improve on-time delivery. But first, we step back and define both intelligence and artificial intelligence. 

Defining intelligence

Defining intelligence has been a long-standing challenge, with various perspectives offered. A widely-accepted definition, which broadly aligns with the context of semiconductor applications, is as follows:

‍Intelligence is the ability to accomplish complex goals. 

As suggested by Max Tegmark [9], intelligence is not universal but depends on the defined goal. As such, there are many possible types of intelligence. Extending this concept further, intelligence can be characterised according to the following features.

Goal type: Intelligence can be technical (problem-solving), social (interaction), or creative (idea generation).

Skill level: This is typically categorised as below/equivalent/super-human level. This determines whether we aim to match the performance of a human or surpass it.

Scope: Narrow intelligence specialises in a specific task, while broad intelligence encompasses a wide range of tasks like human intelligence.

Autonomy: Intelligence can operate with varying degrees of independence, from human-guided to fully autonomous.

In semiconductor scheduling, super-human performance level is necessary to sift through billions or even trillions of candidate solutions to derive optimal decisions, whilst adhering to complex constraints. Focusing on the narrow scope of scheduling allows the system to specialise, thereby optimizing its performance for these specific requirements. The technical nature of the task calls for a solution that exploits the strictly technical aspects to achieve superhuman performance. Finally, a system with high autonomy and no need for human intervention is desired in such a dynamic environment. 

Three important facets of Artificial Intelligence

AI involves creating models and machines that mimic human intelligence, including learning, reasoning, and decision-making. 

Learning is an important aspect of AI, relying on a model’s ability to iteratively refine its internal parameters until it can accurately capture underlying patterns. Machine Learning is the cornerstone approach for learning from data and techniques in this category can range from simple models like Linear Regression to complex Deep Learning networks. 

Reasoning involves drawing inferences based on established rules and facts, mimicking the human ability to logically connect information. It can aid in tasks like medical diagnosis (See the generative LLM AI from Google Med-Palm 2) or legal case analysis.  

Decision-making encompasses action exploration and problem-solving. Action exploration deals with determining actions through interaction with an environment, which can vary from well-defined scenarios, like a chess game, to unstructured situations, like driving a car. Problem-solving, on the other hand, focuses on finding solutions to clearly defined problems with specific objectives and constraints. This can involve simple tasks like sorting or more intricate challenges such as route planning, resource allocation, and scheduling. Optimization and mathematical programming are often employed in these contexts.

Five Crucial Factors When Selecting AI for Production Scheduling

Production scheduling involves making optimal choices to coordinate resources, tasks, and time to meet production goals. It requires handling well-defined parameters and constraints, along with specific objectives like maximising throughput or achieving on-time delivery. As such, it is best suited to rigorous and well-structured AI methods that focus on optimal and feasible decision-making such as mathematical programming. 

Nevertheless, good production scheduling can involve some aspects of learning and reasoning as well. Learning can be useful when some of the parameters are not well defined or static. For example, estimating transfer times between different locations of a fab may depend on various parameters, necessitating the use of a prediction model that has learned from past data. In terms of reasoning, a good decision-making approach should allow some degree of introspection from the user. Contrary to black box approaches, such as deep neural networks, mathematically formal methods such as Mixed Integer Linear Programming (MILP) enable transparency and explainability.

Choosing the right AI technique for production scheduling in semiconductor manufacturing involves navigating the intricate balance among five crucial characteristics, each vital in this high-stakes field:

Optimality refers to the ability of an AI technique to reach and prove that the true optimal solution has been found. In a complex environment such as a semiconductor fab, where small improvements can have significant cost or time implications, optimality is of paramount importance. 

Feasibility is about ensuring that the solution found truly abides by the constraints of the problem. Semiconductor fabs are bounded by many constraints, including machine capacity, human resources, and time windows. An AI solution must respect these constraints while optimizing the schedule. 

Speed is crucial as it directly impacts the responsiveness of the system. Semiconductor manufacturing is a dynamic environment with constantly changing states. Therefore, the selected AI technique must be able to provide fast and accurate solutions to adapt to these changing conditions. 

Explainability refers to the ability of an AI technique to provide insights into how it arrived at a given solution. In a high-stakes environment like a semiconductor fab, explainability helps build trust in the system, enables troubleshooting, and allows for more effective human-AI collaboration.

Flexibility refers to the technique’s applicability across a wide range of possible scenarios and system changes. This attribute highlights the capability of an AI method to be fully autonomous and require  minimal human supervision and intervention. Within the context of a semiconductor plant, this quality is indispensable, especially as complexity grows and specialised personnel are spread thinner across other functions. 

Different AI techniques fare differently on these dimensions. Rule-based systems offer high explainability and feasibility but may lack optimality, especially in complex scenarios. Unforeseen changes in a fab’s state may require rule adjustments or even entirely new ones, affecting flexibility. Heuristic approaches can provide acceptable solutions quickly, but typically cannot provide optimality or feasibility guarantees. Reinforcement learning can potentially offer high levels of optimality and speed, but at the cost of explainability, the risk of infeasibility, and the need for extensive tuning. 

In contrast, mathematical programming techniques, such as MILP, can offer an excellent balance. They provide guaranteed feasibility, while the distance to true optimality can be easily computed. They offer explainability in terms of how decisions are made based on the objective function and constraints. Although computational complexity can be an issue, they can greatly benefit from advanced decomposition methods, and are well complemented by heuristic methods [10].

In the context of semiconductor fab scheduling, where feasibility, optimality, and explainability are particularly important, mathematical programming techniques can be a superior choice for AI implementation. Their deterministic nature and the rigour of their mathematical foundations make them a highly reliable and robust choice for such high-stakes, complex operational problems.

Going beyond with AI
‍

Today, AI in semiconductor manufacturing stands at a critical point. With the increasing complexity of semiconductor processes and the escalating demand for efficiency and quality, the need for effective AI solutions has never been greater. As evidenced in many large companies’ roadmaps, AI is regarded as a key enabling technology of the future [11]. Companies that do not devote resources to a comprehensive AI strategy risk being left behind.

As we delve deeper into the era of AI-driven manufacturing, the nuanced roles of different AI techniques will become more and more apparent. Machine learning approaches bring novel capabilities for learning and predicting from data: yield improvement and predictive maintenance are very promising paths. When it comes to autonomously and reliably scheduling and planning operations in a fab, an exact optimization approach, such as MILP, becomes the key to unlocking peak performance.
 

‍Authors:
Ioannis Konstantelos, Principal Optimization Engineer at Flexciton
Dennis Xenos, CTO and Cofounder at Flexciton

‍

‍

References

[1] McKinsey & Company, Scaling AI in the sector that enables it: Lessons for semiconductor-device makers, April 2021. Link

[2] Mönch, L., Uzsoy, R. and Fowler, J.W., 2018. A survey of semiconductor supply chain models part I: semiconductor supply chains, strategic network design, and supply chain simulation. International Journal of Production Research, 56(13), pp.4524-4545.

[3] Lee, T.E., Kim, H.J. and Yu, T.S., 2023. Semiconductor manufacturing automation. In Springer Handbook of Automation (pp. 841-863). Cham: Springer International Publishing.

[4] Mehrdad Mohammadi, Stephane Dauzeres-Peres, Claude Yugma, Maryam Karimi-Mamaghan, 2020, A queue-based aggregation approach for performance evaluation of a production system with AMHS, Computers & Operations Research, Vol. 115, 104838, https://doi.org/10.1016/j.cor.2019.104838

[5] Çınar, Z.M., Abdussalam Nuhu, A., Zeeshan, Q., Korhan, O., Asmael, M. and Safaei, B., 2020. Machine learning in predictive maintenance towards sustainable smart manufacturing in industry 4.0. Sustainability, 12(19), p.8211.

[6] Ishida, T., Nitta, I., Fukuda, D. and Kanazawa, Y., 2019, March. Deep learning-based wafer-map failure pattern recognition framework. In 20th International Symposium on Quality Electronic Design (ISQED) (pp. 291-297). IEEE.

[7] Dreyfus, P.A., Psarommatis, F., May, G. and Kiritsis, D., 2022. Virtual metrology as an approach for product quality estimation in Industry 4.0: a systematic review and integrative conceptual framework. International Journal of Production Research, 60(2), pp.742-765.

[8] Moyne, J., Samantaray, J. and Armacost, M., 2016. Big data capabilities applied to semiconductor manufacturing advanced process control. IEEE transactions on semiconductor manufacturing, 29(4), pp.283-291.

[9] Max Tegmark, Life 3.0, Being human in the age of Artificial Intelligence, 2018 

[10] S. Elaoud, R. Williamson, B. E. Sanli and D. Xenos, "Multi-Objective Parallel Batch Scheduling In Wafer Fabs With Job Timelink Constraints," 2021 Winter Simulation Conference (WSC), Phoenix, AZ, USA, 2021, pp. 1-11, doi: 10.1109/WSC52266.2021.9715465.

[11] Bosch, Humans and machines team up in the factory of the future, October 2021. Link

‍

In Part 1 of this blog, we focused on use cases where lots are scheduled on tools and how advanced scheduling gives users the ability to optimize for future decisions as well as real-time. When we say "advanced," we are referring to autonomous, optimization-based solutions. Our emphasis was primarily on how scheduling can enhance productivity in a fab today. In Part 2, however, we’ll delve further into its potential for fabs in the not too distant future.

Previously, I discussed how task lists are typically associated with human workers. However, it is worth noting that task lists can also be applied to automated systems such as automated guided vehicles (AGVs) and automated material handling systems (AMHS) with the use of an advanced scheduler. With task lists, an advanced scheduler can not only determine which lot is assigned to which tool and when, but also which operator – or robot – will be serving the tool. There’s a whole set of new opportunities that arise with that, as humans and robots, just like tools, have a limited capacity that can be optimally utilised. It’s clear then that the possibilities for advanced scheduling go beyond the stand-alone Industry 4.0 applications and have the potential to integrate vast amounts of fab data into a holistic system. 

One of the use cases of such a holistic system is described later on in this blog as a type of ‘digital twin’, but the capabilities of an advanced scheduling system go beyond that. With a digital twin concept, the human is still very much inside the cockpit. An advanced scheduling system, on the other hand, is more like an autopilot, augmenting the capabilities of other systems and taking control of manufacturing decisions when necessary. As such, advanced scheduling is a cornerstone of the so-called ‘smart factory’. Let’s try to understand the huge array of benefits it can bring. First, we’ll cover a couple of use cases that can benefit the manufacturers. Second, we’ll share some thoughts on how advanced scheduling aligns with the idea behind Industry 5.0 and how the technology can serve ourselves as humans.

1. WIP Transport Scheduling

Once a lot is intelligently scheduled, we know when to process it and on which tool. The lot can be transported to that tool’s specific staging rack just before it gets processed. It enables fabs to eliminate waste by optimizing transport capacity, which removes the likelihood of a lot being transported at half capacity only for it to wait in queue. Transport scheduling also enables splitting logistics and processing workflows; some workers focus on keeping the tools running, others focus on getting the lots to the tools in time. Multi-cleanroom fabs will make better use of their capacity in areas that for logistical reasons are not preferred. Which means no more remote idle machines waiting for a lot that doesn’t arrive.

2. Dynamic Capacity Models

With better control of lot processing, intra-fab logistics, and workforce planning, we get a more realistic view on the true capacity of a factory. We call it a dynamic capacity model, resembling the idea of a digital twin of a production plant. A dynamic capacity model better reflects the current state, loading and dynamics in a factory, as opposed to the static capacity models commonly used. Until now in wafer fabs, dynamic capacity models have at best been approximated by what-if scenarios in simulation models, but the potential goes beyond that. When playing around with different scenarios – e.g. when to plan maintenance or shutdowns, which availability increase has the most impact on the whole factory, what’s the effect of frequent product mix changes, what lead times to expect and so on – it should allow factories to better judge the impact of their decisions. Optimization can even help by not only interpreting the outcome, but suggesting the best decision for a fab’s goals.

3. Multi-factory Models & Supply Chain Planning

Eventually, dynamic capacity models could scale to corporate level in multi-factory models. Further up, these models could feed into supply chain planning software. During the supply chain crisis, it was striking to see how disconnected sales and operations planning cycles in semiconductors were from the actual operational challenges of factories. Part of it was because of models that don’t properly comprehend the actual situation the factory was in. Fabs were treated as black boxes with a simple input and output signal, but just because you have promised your customers a sooner delivery date, it doesn’t mean it will happen automatically. You need a driver towards that new target, and that’s where advanced scheduling software helps, by optimizing towards shorter lead times. Its integration into dynamic capacity models and supply chain planning software would lead to more reliable input for inventory and order fulfilment optimization engines. This translates into lower inventory costs and better delivery performance of a company.

Eventually, we want technology to help us overcome the challenges we face as humans. From what has been written so far, this blog might give the impression that this technology is primarily serving profitability. But becoming a smart factory doesn’t necessarily contradict with a human-centric approach. Industry 5.0 is the theoretical concept that’s been introduced for that. It counters the illusion that the future of manufacturing is one in which humans play a minor role. Instead, we should embrace both the capabilities of new technologies, as well as those of humans and find synergies to make the best of both worlds. While Industry 4.0 can do a great job in automating repetitive tasks or making sense out of masses of data, humans have the advantage of better interpretation of context, require fewer data points to understand, and can make value trade-offs. Humans will not miraculously disappear from the factory shop floor, so we’ll benefit from thinking about how these advanced technologies can harmoniously coexist with people and yield mutually beneficial outcomes.

4. Staff Workload Balancing

The obvious fear with advanced scheduling is that operators and technicians will turn into de facto robots, where only adherence is of importance when aiming to get more out of the workforce. Let’s turn that thought up-side-down: what if the same work could be better distributed amongst the team by offloading peaks to underloaded co-workers? Advanced scheduling can better predict and hence properly distribute work aligned with an individual's availability and level of training. Also the workflow itself - the number and order of actions to perform - can be streamlined to lower physical and mental workload.

5. Training, Evaluation & Individual Productivity

With detailed production schedules, any lack of staff or training becomes directly visible and quantifiable. Hiring and training programs could become more timely and data-driven, just as annual evaluations will become less subject to biases of the manager. Even on-the-spot productivity can be monitored and optimised. This may sound like a “Big Brother” concept, but compare it with the advancement of sports analytics and medicine in the last decade. Professional athletes don’t complain about data integrity and privacy issues, because (1) it’s part of their job and (2) it helps them in what they want to achieve. If athletes ignore their data, they simply don’t reach the top anymore. Similarly, the fourth and fifth industrial revolution will bring staffing to higher levels of productivity, not because they are squeezed out more, but because the data will reveal where there’s room for improvement or when a red line is about to be crossed.

Given the increasing scale and complexity described above, significant computational power and data storage capabilities will be necessary. This makes it likely that cloud-based technology will be adopted to facilitate the transition to smart factories. Although many fabs are currently far from achieving smart factory status, it is clear that the industry is moving in this direction. Therefore, factory managers must acknowledge that the transition to becoming a smart factory is not just a concern for the future and must be implemented within a realistic timeframe. The foundations for this transition, including employee readiness, are already being established today. And given the use cases discussed, let there be no doubt that advanced scheduling will play an integral part in the next generation of wafer fabs.

Author: Ben Van Damme, Industrial Engineer and Business Consultant

One of the consequences of the pandemic has been an incentive to deglobalise, as regions suffered from the issues with supply chains and geopolitical dependencies. Significant delivery issues in the chip industry – and in particular wafer manufacturing – have had a negative impact on the global economy. However, onshoring this high technology industry will also bring its own challenges. Expertise and cost efficiency to name a couple. Zooming in a bit closer on so-called wafer fabs, we can distinguish two types of factories. The legacy and smaller fabs serving niche markets with older technology nodes, and the cutting-edge giga-factories, recently built or in the making. Both types have different problems to tackle, but one key component of their roadmap could be surprisingly similar.

The newest fabs have well integrated automated systems, but operating them efficiently on such a scale is a challenge of its own. The older factories have the downside of being less automated but they realise the need to become more efficient in energy consumption, labour cost and capacity utilisation. In both situations, digital transformation is coming to the rescue. Industry 4.0 is no longer a buzzword, it has become a matter of regional technological sovereignty. 

The fundamental building block of Industry 4.0 is data; an asset which is present in abundance in wafer fabs. So what is preventing these factories from levelling up? The answer is simple, the solution is not: complexity. It’s an inherent part of wafer manufacturing, stemming from; increasingly high numbers of process steps, job shop factory types, re-entrant flows, product diversity, sensitivity to quality issues and so on. 

The problem with complex systems is that there’s so much variability and interaction, it's hard to get actionable insights from data. Instead of accepting the stochastic and complex nature of the fab, factories can better control it by using advanced production scheduling to understand in which order lots get processed, on which tool and – the most important difference when compared with common rules-based approaches – when they get processed. To begin, this can be employed in certain bottleneck areas and then once you do it for the entire factory, you get a holistic picture of what is going to happen. Sounds great, doesn’t it? But how exactly will this benefit your fab? To explain, let’s place production scheduling in a couple of recognisable use cases. 

‍

1. Lot-Tool Assignments

Wafer manufacturing has complicated recipe-tool qualification matrices within a group of tools that perform similar processes. The weaker tools can process fewer recipes than the stronger ones. We want to avoid stronger tools “stealing” lots away from the weaker tools, because it leaves fewer lots for the weaker tools to process, therefore wasting capacity. The same is true for faster and slower tools: while faster tools are preferred, pushing all the WIP through the faster tools leaves the slower tools under utilised. Advanced schedulers allow for better anticipation of incoming WIP and superior use of available capacity for weak and slow tools. The bigger and more complex the matrix grows, the harder it is to find the optimal processing of WIP. On top of the scheduling itself, mathematical programming helps to optimize lot-to-tool assignments over time. This results in a capacity booster, similar to putting a turbocharger on an engine: it’s the same engine, but with more power.

2. Reducing Timelink Violations

Process steps with timelinks are common in wafer manufacturing to control the maximum amount of time a wafer spends between two or more process steps. If a timelink is violated, the wafer requires rework – or worse still, scrappage. A system that avoids timelink violations requires the ability to intelligently plan into the future. And that’s exactly what an advanced scheduler does. It has been proven to drastically reduce timelink violations, even in the most complex of scenarios. 

3. Improving Batching Efficiency 

Batching is a complex decision making process since it involves an estimate of lot arrivals and how waiting longer trades off with running smaller batches. Predicting lot arrivals is difficult in such a complex environment, and trading off wait time against batch efficiency is even harder because the costs and gains are not always clear. Determining and automating this process is well within an advanced scheduler’s remit. Once the algorithm is tuned, it makes the most efficient decision, and perhaps even more importantly: it generates consistent output. 

4. Optimizing Changeover Decisions 

Another use case related to the problem of lot arrivals is the problem of changeover decisions. One toolset with different machine setups can serve multiple different toolsets down the line. A bit like a waiter in a restaurant serving multiple tables. Waiters have to make sure no table is without food or drink, and to do that, they visit the tables regularly to ask for any orders. But for machines, you can’t switch the setup too often because it only increases non-productive time. Preferably, you also plan setup changeovers at a time when planned or predicted downtime for the machine occurs, to reduce downtime variability. To put it simply, it’s a decision on when to switch over from the type A process to the type B process on a tool. An advanced scheduler can solve that equation, finding the optimal point in time. Schedulers are better at this than human reasoning or rule-based logic, as solving to a time dimension is what they are designed for.

5. Flow Control and Line Balance

Line balancing is – even for experienced manufacturing engineers – difficult to grasp. One can intuitively understand what it means, but how do you define “balanced” in the first place? Even if you can, it is absolutely beyond the capabilities of a human brain to manually and continuously make decisions that control it. And once it’s out of balance, to recover it. Again, considering the time dimension is a crucial aspect of what advanced schedulers offer, which enables them to recover faster from unforeseen circumstances and maintain better risk-control for generating continuous output.

6. Operator Task Lists

As opposed to dispatch lists that only tell the order in which to process lots, advanced schedulers can also tell when a lot is supposed to start and finish processing on a tool. Combine that information with which operators are serving which tools, and you can move away from tool-centric dispatch lists towards operator-centric task lists. With a handheld device, that could even allow you to send push notifications when urgent intervention is needed. It can reduce idle time on tools that have no available operator. Even more so, it can allow for an entire rethink of the workflows operators are used to. 

‍

‍

So far in this blog, we’ve focused on scheduling use cases where lots are scheduled on tools, leading to higher throughput on tools, toolsets or the entire factory. All these use cases can also be addressed by improving some rule-based dispatching strategies, but what advanced scheduling offers is the ability to optimize for future decisions rather than just real-time. With that comes better visibility on what will happen in the factory, and it also leaves opportunities for re-organising workflow and freeing up resources. In part 2 of this blog, we’ll begin to look at the future and what could happen when we integrate even further. Enter, Industry 5.0. 

Author: Ben Van Damme, Industrial Engineer and Business Consultant

‍

Part 2 is now live. Click to read.

‍

‍

This article draws from the contents of a paper presented at Winter Simulation Conference 2022, titled: “Fab-Wide Scheduling of Semiconductor Plants: A Large-Scale Industrial Deployment Case Study”.

An Introduction to Fab-Wide Scheduling

The semiconductor industry is one of the largest and most complex industries in the world. The critical factors in semiconductor manufacturing are the ability to rapidly develop and test novel technologies, improve manufacturing processes to reduce rework and waste, as well as meet production targets in terms of prescribed volumes and due dates. In this context, high quality scheduling is of paramount importance.

Due to the long cycle times, where a wafer is processed over a span of months, decision-making in semiconductor fabrication plants (fabs) is typically framed as a two-level problem. On one hand, global scheduling (or fab-wide) is tasked with the strategic management of factory assets while considering all work-in-progress, incoming and outgoing flows across the fab, expected resource availability and other constraints. On the other hand, local (or toolset-level) scheduling focuses on the operation of individual work centres. It is typically tasked with identifying the best immediate dispatch decisions i.e. which jobs waiting for dispatch should be assigned to which available machine.

Most development efforts to date have focused on the shorter time frame dispatch decisions i.e. local scheduling. This is a more manageable problem since there is little look-ahead and the scope is limited to a single or a few toolsets. Despite numerous research efforts, to date there has not been a published case study of a fab-wide scheduler successfully deployed in a large semiconductor manufacturing facility. Nevertheless, the potential for improvement at the fab-wide level is tremendous; there are numerous opportunities to improve throughout and have a step change in performance. For example:

• Bottlenecks occur due to repetition of process loops, high-cost machines with low capacity, and other physical or operational constraints. To manage them, a strategic approach is needed that looks at the bigger picture and avoids early dispatch of wafers that will end up in a bottleneck area. 
• WIP flow control mechanisms (kanbans) are important for quality control but can block high-priority wafers. Fab-wide scheduling can greatly improve this aspect of operation. 
• Timelinks (also known as timeloop, time lag, or qtime constraints) are challenging because they define the minimum or maximum amount of time between two or more consecutive process steps, leading to a conundrum of keeping downstream machines idle or not. Fab-wide scheduling can greatly assist by accurately predicting arrival times and deciding when to trigger timelinked lots.

Methodology

The scheduling framework proposed in this blog is hierarchical and consists of two main components which run independently and at different frequencies — the Toolset Scheduler (TS) and Fab-Wide Scheduler (FWS). 

The Toolset Scheduler considers the currently in-process and/or upcoming process step of all wafers in the cluster.

FWS takes a view of the entire fab at once and considers multiple future steps for each wafer. It focuses on improving schedule quality by considering the flow of wafers through the fab, something the toolset scheduler cannot do due to its singlestep, toolset-level nature. The main purpose is to redirect flow through the fab and thereby improve flow linearity, reduce bottlenecks, improve WIP flow control management, and reduce timelink violations. Our FWS approach achieves this by predicting wait/cycle times for multiple future steps, analysing those predicted wait/cycle times with respect to the different areas of potential improvement, and re-prioritising wafer steps in a way that guarantees improved (weighted) cycle times. In brief, FWS combines two main elements: (i) an operational module that captures in full detail all relevant constraints e.g. detailed process time modelling, machine maintenance, shift changes, dynamic batching constraints, kanbans etc. (ii) a search module that identifies beneficial priority changes given the evolving fab conditions and state features.

FWS communicates with the toolset schedulers via priority weights (and some other predicted timing information) for individual steps of a wafer, as shown in Figure 2. An advantage of our approach is that, while FWS always schedules all tools in the fab, users can specify which toolsets are subject to guidance; FWS adjusts its search accordingly. This is particularly useful for gradually rolling out FWS in a fab and evaluating its impact. In addition, the guidance strength is controllable - although full guidance is the optimal choice, tuning down guidance allows for a more gradual deployment.

Seagate Deployment

Seagate is a world leader in data storage technology, with more than 40% share of the global Hard Disk Drive (HDD) market. The Springtown facility in Northern Ireland produces around 25% of the total global demand for recording heads, the critical component in a HDD. Flexciton’s FWS / TS scheduling system was trialled in Seagate Springtown between March-May 2022. After successful testing, the system has been operational 24/7 since June 2022; a timeline is shown in Figure 3.

It is important to note that deploying and testing a novel piece of technology in a large factory that runs around the clock presents many practical challenges to be overcome:

• Controllability (scope): important to ensure that the new development is deployed in a controlled manner. The FWS-TS guidance scheme allows for localised trials, where focus can be placed on problematic areas and gradually increase scope.
• Controllability (magnitude): it is useful to only focus on cases with obvious merit first. This is achieved by controlling guidance strength. 
• Explainability: important to be able to detect and reason about the changes. This is achieved by a combination of UI features and support tools which have been designed to give operators and managers situational awareness.
  ‍

Results and Learnings

Quantifying the benefit of an alternative scheduling approach remains a challenging task. When deployed in a real plant, traditional A/B testing between pre and post-deployment suffer from (i) dynamic fab conditions (ii) an ever-changing product mix and (iii) evolving capabilities of the fab e.g. increased/decreased labour capacity and new tool commissioning/decommissioning.

As such, it was decided to look at the impact from different angles - a statistically significant impact would be expected to result in a substantial shift in numerous business  processes and metrics. In particular, three different aspects were examined.

• Deep dives on specific toolsets and metrics.
• Comparison against internal simulation and planning tools. 
• Observing the impact on manual interventions.
  ‍

Notably, all three approaches indicated a change in fab performance between pre and post-deployment; more details will be shared in future articles. In the Winter Sim Conference paper presented in December 2022, we focused on the latter point; A proxy we can use for this benefit is the volume of ad hoc control flow rules activated/deactivated in the fab. Every day, specialists have to define numerous, in some cases even hundreds, of ad hoc control flow rules to better manage operations given the prevalent conditions. For example, setting a ”hard down” rule, where lots are manually placed on hold so as not to continue to a downstream bottleneck. In Figure 5, we show the number of ad hoc operational rules implemented in the Seagate Springtown fab between weeks 2 and 26 of the year 2022 (i.e. from early January until late June). As can be seen in the final weeks, the number of ad hoc rule transactions averaged less than 150 per week, a decrease of over 300% compared to the pre-deployment period. This is strong evidence that FWS deployment reduced massively manual interventions required to effectively control flows within the fab.

Conclusions

The main takeaway of the Winter Sim paper is that the increased horizon look-ahead and global nature of FWS presents numerous opportunities for a step change in factory KPIs. The Flexciton FWS was successfully trialled at Seagate Springtown over 3 months in 2022 and has been fully enabled across the fab since June 2022. It resulted in a radical decrease of interventions previously used to manually control wafer flows. Further analysis suggests that Flexciton’s TS and FWS schedulers have achieved substantial improvements in throughput and cycle times.

‍

Author: Ioannis Konstantelos, Principal Engineer

Introduction

The first integrated circuits were invented by Texas Instruments and Fairchild Semiconductor in 1959. Today, semiconductor manufacturing is a $600 billion dollar industry and microchips are ubiquitous and impact our lives in ever increasing ways. To achieve such astonishing growth, academics and industry have had to constantly innovate, researching new production technologies. While much has been said about Moore's law and the push towards higher and higher transistor densities, the innovations made in how the billion dollar factories producing these chips are run have received less attention. This article focuses on innovations in scheduling: algorithms which assign lots to machines, decide in which order they should run, and ensure any required secondary resources (e.g. reticles) are available. These decisions can significantly impact the throughput and efficiency of wafer fabs.

Many innovative technologies in scheduling were first proposed by researchers and have, over time, been adapted in manufacturing. They include:

• Dispatching: rule-based systems for deciding which lot to run next on a tool 
• Optimization-based scheduling: mathematical techniques like mixed integer programming and constraint programming which can generate optimal machine assignments, sequencing, and more for entire toolsets or areas of the fab, improving fab-wide objectives like cycle-time or cost
• Simulation: computer models of the manufacturing process which are often used to run what-if analysis, evaluate performance, and aid decision making

From dispatching to mathematical programming

Early academic research on dispatching rules dates back to the 1980s. Authors at the time already highlighted the significant impact scheduling can have on semiconductor manufacturing. They experimented with different types of dispatching rules, ranging from simple first-in-first-out (FIFO) rules to more bespoke rules focused on particular bottleneck tools. Over time, dispatching rules have evolved from fairly simple to increasingly complex. Rule-based dispatching systems quickly became the state-of-the-art in the industry and continue to be popular for several reasons: they can be intuitive and easy to implement, yet allow covering varying requirements. There are, however, also many situations in which dispatching rules may perform poorly: they have no foresight and generally look only at a single tool and therefore often struggle with load balancing between tools. They also struggle with more advanced constraints such as time constraints or auxiliary resources, e.g. reticles in photolithography. More generally, dispatching systems are a mature technology that has been pushed to its limits and is unlikely to lead to significant increases in productivity and yields.

For these reasons, focus has shifted over time to alternative technologies, especially deterministic scheduling based on mixed-integer programming or constraint programming. In the academic literature, these approaches start to increasingly show up around the 1990s. Early contributions focused on analysing the complexity of the wafer fab scheduling problem and solved the resulting optimization problem using heuristic techniques, but slowly moved towards rigorously scheduling single machines, tackling one particular aspect of the problem at a time. Due to the limited scope deterministic techniques could initially tackle, their adoption in industry lagged behind the academic discussion. 

From single machines to fab-wide scheduling

The last twenty years have seen deterministic scheduling techniques mature and schedule larger and more complex fab areas. In the academic literature, authors moved from focusing on single (batching) tools, to entire toolsets or larger areas of the fab including re-entrant flows. They also started including more and more operational constraints such as sequence-dependent setup and processing times, time constraints, or secondary resources such as reticles. In order to achieve this increase in scale and complexity, researchers have applied a large number of optimization techniques, and often combined rigorous mathematical programming methods with heuristic approaches.  Some have used general purpose meta-heuristics, such as genetic algorithms or simulated annealing, while others have developed bespoke heuristics for fab scheduling, such as the shifting bottleneck heuristic. 

As the size of problems optimization-based scheduling techniques could solve grew, the industry started to explore how to adopt these methods in practice. For example, in 2006, IBM announced that it had successfully used a combination of mixed-integer programming and constraint programming to schedule an area of a fab with up to 500 lot-steps and that this had led to a significant reduction in cycle time. Our own technology at Flexciton leverages mathematical optimization and smart decomposition, combined with modern cloud computing, to efficiently schedule entire fabs. One key advantage of using cloud technology is the ability to access huge amounts of computational power. It allows to break down complicated problems and deliver accurate schedules every few minutes, as well as the ability to adapt the solution strategy to the complexity at hand. Additionally, it enables responsive adjustments, as events unravel in real-time, allowing for a truly dynamic approach to scheduling.

Optimization-based scheduling’s trajectory from an academic niche to a high-impact technology has partially been accelerated by two major trends:

• The increasing automation of wafer fabs and availability of process data. Especially automated material handling systems in modern fabs have made scheduling techniques almost a necessity. 
• The success of the semiconductor industry itself. Faster microprocessors have made it possible to solve bigger scheduling problems in less time. 

The process has been accompanied by considerable improvements in productivity, as scheduling is able to overcome many of the downsides of dispatching: it can look ahead in time, balance WIP across tools, and improve fab-wide objectives such as cost or cycle-time. A major advantage of scheduling is that it can both increase yields when demand is high and reduce cost when demand is low. 

When in doubt, simulate.

A discussion of scheduling in wafer fabs would not be complete without a word on simulation models. Simulation models are technically not scheduling algorithms - they require dispatching rules or deterministic scheduling inside them to decide machine assignment and sequencing. But they have been used to evaluate and compare different scheduling approaches from the very beginning. They were also quickly adopted by industry and have, for example, been used by STMicroelectronics to re-prioritise lots and by Infineon to help identify better dispatching rules. The development of highly reliable simulation models could greatly increase their use for performance evaluation and scheduling.

The future

More reliable simulation models are also important in light of recent trends in academic literature, which may provide a glimpse into the future of wafer fab scheduling. Rigid dispatching rules that need to be (re)tuned frequently may soon be replaced by deep reinforcement learning agents which learn dispatching rules that improve overall fab objectives. In some studies, such systems have been shown to perform as well as dispatching systems based on expert knowledge. If and when the industry adopts such techniques on a large scale remains to be seen. Since they require accurate simulation models as training environments, they can be extremely computationally intensive, and their adoption will largely depend on the development of faster training and simulation models. The combination of self-learning dispatching systems, and comprehensive, scalable scheduling models may well hold the key to unlocking unprecedented improvements in fab productivity. 

Flexciton aspires to be the key enabler in this transition, bringing state-of-the-art scheduling technology to the shop floor in a modern, sophisticated, and user-friendly platform unlike anything else on the market. Despite the enormous challenges that come with the scale of this endeavour, the initial results are very encouraging; cloud-based optimization solutions can indeed bring a step change to streamlining wafer fab scheduling while delivering consistent efficiency gains. 

‍

It has been described as the law that rules the modern world, and its effects can be observed in every organisation. I’m referring to Goodhart’s law, named after British economist Charles Goodhart, who wrote the maxim: “Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.”

A common flavour of this effect is described in the following cartoon, based on a possibly apocryphal story of how central planning failed in a nail factory in the Soviet Union.

We have seen (less dramatic) examples of this effect at work in semiconductor wafer fabs. For instance, teams of operators may be measured on the number of lot moves that occur during their shift. In general, more moves per shift correlates with more wafers delivered on time to customers. However, this relationship breaks down if operators ‘game the system’ by loading batch tools with small batches at the end of the shift, thus wringing out a few extra moves in their shift, but hobbling the next shift.

Memorable though such examples are, they give the impression that Goodhart’s Law relies on people being uninterested in the ultimate goal that their organisation is pursuing. However, apathy is not usually the driving factor in Goodhart’s law; whenever lack of information, limited computational power or even an inability to concisely express our true preferences leads us to substitute a proxy metric for our true goal, the law is bound to rear its head. Former Intel CEO Andy Grove described the effect of such surrogate indicators as like “riding a bicycle: you will probably steer where you are looking”; and if where you’re looking isn’t perfectly correlated with the road ahead, you can expect a wobbly ride!

The intricacies of tools with multiple load ports

For a more subtle example of where using an imperfect measure as a target can lead to suboptimalities when scheduling a wafer fab, we were inspired by a post on the excellent Factory Physics and Automation blog looking at the relationship between load port utilisation and cycle time. In our experience, we have seen load port utilisation of a tool used as a target when designing both operator workflows and dispatching rules.

First, some quick definitions. Many tools in a fab have multiple ‘load ports’ where lots can be inserted into the tool, but then a limited chamber capacity so that, for instance, only one wafer can be processed in the chamber at the same time.

Consider the machine in Fig. 1 with three chambers and two load ports. Lots can be loaded in either load port, but then each wafer in the lot has to move through Chambers A, B and C one at a time. This means wafers may have to queue inside the tool, if the next chamber they need is still processing.  Lots must be unloaded at the same load port in which they were inserted. Suppose it takes each chamber 10 minutes to process a wafer, and we want to process two lots each consisting of three wafers. If we were only allowed to use a single load port, we would have to wait for the first lot to move through all three chambers and exit at the same load port before we can start processing the second lot. Fig.2 shows that for a simple model (that ignores transfer time between chambers), the second lot will have to wait 50 minutes before it can start processing.

If however, an operator loads both batches into the two load ports at the same time (Fig. 3), the machine will pick up the first wafer of the second lot as soon as the first lot has finished processing in chamber A. Thus the second lot will only need to wait 30 minutes.

Therefore, for a given level of WIP at a tool, we can expect higher load port utilisation to be correlated with reduced waiting and therefore improved cycle time.

Indeed, in cases where a wafer cannot be unloaded from a tool until all the wafers in the same lot are also ready to be unloaded (a common workflow), it can actually make sense to split lots before a chamber tool. For instance, if we have a lot of 6 wafers before the tool (see Fig. 1) – loading all the wafers as a single lot in a load port – it will take 80 minutes for all 6 wafers to move through the three chambers until we can unload the lot. If however, we split the original lot into two lots of three and load them into both load ports (as in Fig. 3), then the first lot can be unloaded after just 50 minutes, and potentially continue to its next step earlier.

How directly targeting load port utilisation can harm cycle time

As predicted by Goodhart’s Law, the correlation between load port utilisation and fab cycle time breaks down once we try to optimize directly for load port utilisation. This breakdown is particularly stark on photolithography tools, where process steps rely on a critical secondary resource: reticles. Reticles (also called photomasks) act like stencils in the expose step of a photolithography process, patterning the wafer with the desired features.  In most photo tools, reticles must be loaded onto the tool in containers, called pods, before the lots that require them can be loaded onto the machine. Therefore, if a lot is inserted into a load port early, the wafers could just be waiting inside the machine. Moreover, this also requires loading a reticle into the machine when it could have a more productive use elsewhere.

For a simple example, consider a toolset consisting of two of the tools from Fig. 1 (we can imagine chambers A, B and C are performing coat, expose and develop operations respectively).

Suppose we have just loaded a 3 wafer lot onto tool 1. The other load port of tool 1 remains free. Meanwhile on tool 2, both load ports are utilised, but there are only two wafers yet to be processed in Chamber A.

A lot (lot X) that requires a special reticle (of which only one exists) arrives. Due to a lot-level restriction, lot X can only run on tool 1. This sort of restriction is particularly common in photolithography where running consecutive photo layers through the same tool (even if there are multiple tools qualified for the operation) can reduce product variability caused by idiosyncratic aspects of the lensing to a particular tool (this is sometimes known as a ‘lot-to lens’ dedication).The operators on this toolset abide by the following rule for dispatching lots:

Rule 1: If a load port and the required reticle are available, load the reticle and the lot onto the tool.

Since tool 1 has a load port available, the operator immediately loads the reticle onto the machine, and puts lot X into the load port.

Ten minutes later, lot Y arrives at the toolset, also requiring the same reticle, and with a lot-level restriction forcing it to run on tool 2. Since the reticle is already loaded on tool 1, lot Y cannot be dispatched until lot X has finished processing and the reticle has been moved from tool 1 to tool 2. Assume, for the purpose of simplicity, the reticle moves instantaneously, both lots will have finished processing in 130 minutes time (see Fig. 4).

Imagine, however, the operators adopted the following workflow:

Rule 2:  If a load port and the required reticle are available and the tool can begin processing immediately (i.e. Chamber A is free), load the reticle and the lot onto the tool

In this case, lot X will not be immediately loaded onto tool 1, since Chamber A is initially occupied. After only 20 minutes though, lot Y can be loaded onto tool 2, to finish processing 50 minutes later, at which the reticle can be moved and lot X can start on tool 1. Thus, after just 120 minutes (as opposed to the 130 minutes under Rule 1), both lot X and lot Y will have finished processing. Therefore, we can see that by adopting rule 2, the cycle time, and hence the throughput of the toolset can be improved.

In our experience of wafer fabs, we often see workflows akin to Rule 1, wherein operators fill the load ports of photo tools as soon as they are free, thus forfeiting the opportunity to use reticles earlier on different tools. Adopting a workflow like Rule  2, however, is more difficult since it requires operators to have foreknowledge of when the tool will be ready to process a new lot, and reacting promptly to load the tool at precisely this time.  In practice, particularly when operator availability is limited, you will risk increasing wait time because you leave the tool under utilised if you fail to load a lot as soon as a machine becomes available.

Using advanced optimization to handle Goodhart's Law

Flexciton’s scheduler can help to alleviate this problem by employing advanced optimization technology. It can predict when lots will arrive at the photo toolset and which reticles they will require, and then jointly schedule the reticles and lots on the toolset to obtain an optimized schedule. The knowledge of future arrivals crucially allows us to identify cases where loading a reticle onto a machine now is suboptimal, since a lot will soon arrive at another tool that can make use of the reticle sooner or that simply has a higher priority. Thus, following a Flexciton schedule, operators can dispatch to load ports when they become available, with minimal risk of harming cycle time due to locking in reticles prematurely.

However, we still are not immune to the curse of Goodhart’s Law. The cycle time of an optimized schedule is itself only a proxy for what we actually care about: producing more high quality wafers at a low cost per wafer. Over-optimizing for cycle time may lead to a solution with so many loads and unloads that the labour cost of running fab becomes prohibitive. Or, as described in one of our previous blog posts, the solution may require moving reticles so frequently between tools that we increase the chance of a costly breakage.

To solve this, we apply a technique suggested by Andy Grove himself: we use pairing indicators. Combining indicators, where one has an effect counter to the other, avoids the trap of optimizing one at the expense of another. This is why we typically pair cycle time with the number of batches (to account for limited operator availability) or the number of reticle moves (to keep the risk of reticle damage low), thus mitigating the perils of Goodhart’s Law.

‍

The diffusion area is particularly important to the smooth operation of a wafer fab. Not only does it receive raw wafers at the very beginning of the fabrication process but it also interacts with many other areas of the fab.

The challenge in scheduling diffusion area lies in the particularities involved in its operation:

• Re-entrant flows: Furnaces are loaded with wafers that may have already been processed by other furnaces or wet benches.
• Batching machine: Several lots of 1-25 wafers each can be processed together in one batch if they run the same recipe.
• Time constraints: The diffusion area has timelinks that determine the maximum time a wafer has to move to a subsequent tool to avoid rework or scrappage.
• Dummy wafers: Used in furnaces, for example, to fill out a lot when a full lot is required or to protect the most exposed ends of a lot to ensure uniformity.

Balancing very long fixed processing times on batching tools with the features mentioned above makes it exceptionally tricky to get solid production KPIs on diffusion furnaces. Currently, fabs often resort to using simplistic "minimum batch size" dispatch rules that try to balance building full batches (to maximise the utilisation of the tool) with queue time and the risk of violating a timelink constraint.

As a result of these characteristics, it's very common for diffusion areas to become a bottleneck if not managed correctly – negatively impacting the production KPIs of the rest of the fab.

This is what prompts the exploration of more novel scheduling methods, such as the one we'll be discussing in this article.

Case Study: Job Scheduling of Diffusion Furnaces

To explore the various ways to schedule diffusion areas, we review the paper “Job scheduling of diffusion furnaces in semiconductor fabrication facilities” by Wu et al. (2021) that describes a new scheduling system that was deployed live in a 200mm GlobalFoundries wafer fab.

Fab Characteristics and the Need for Change

The fab that the system was implemented into consisted of the following attributes; approximately 300 products, 500 recipes, and 4500 lots daily at the diffusion area which is host to more than 90 furnaces.

The approach was designed to build schedules aiming to maximise the weighted number of moves. The weights were based on the product of the wafer and the stage of production for which moves were being calculated.

Schedules were planned by 6 operators several times a day, taking up to 6 hours a day per operator on average. The quality of schedules are also impacted by the judgement and experience of operators which led to suboptimal decisions and lower efficiency.

The Approach to Scheduling

The heuristic model used in the system took about nine months to be built, whilst the system implementation took one year and a half, with the majority of time being spent on clarifying user requirements and collecting data.

The problem was addressed with techniques called Dynamic Programming and Genetic Algorithm:

Dynamic Programming consists of breaking down a large problem that contains many possible solutions into several sequential sub-problems that are easier to solve. Each of these subproblems is solved one at a time, such that each solution feeds into the next problem.

Each one of these sub-problems is then solved using a modified version of the Genetic Algorithm, a meta-heuristic procedure commonly used for large optimization problems.

Results

After implementing live in the fab, average daily weighted moves per tool improved by 4.1% in the first two months of trials when compared to the 2 months before deployment. When tested offline and compared to historical data, the approach increased the number of moves by 23.4% and the average batch size by 4.1% while reducing tool idling by 62.8%. The authors argue the fab was short of staff, subject to varying demand and product mix over time, and with operators still not fully adhering to the new schedules.

It is also expected that, by exploring the full potential of the system, cycle time can be reduced by 1.8 days and that an increase of eleven thousand moves can be achieved, leading to an estimated financial saving of $2M USD per year.

Flexciton’s View

A lot of the academic literature on scheduling furnaces tend to omit some rather critical details such as missing constraints, only being tested on small test datasets, or they are prohibitively slow in live environments.

The reviewed approach stands out by addressing these issues and successfully implementing a complex scheduling system in a fab that brings measurable improvements to the number of moves, batch size and tool idleness. The model accounts for many relevant details such as preventive maintenance, lots with tool dedications at certain steps and different lot priorities.

Nevertheless, as specialists in scheduling, we have spotted weaknesses in the approach where we believe there are opportunities to make it even more robust and versatile, whilst delivering even better results:

1. Schedule updates every 40 minutes: unexpected events (e.g., machine downtime) can take longer than schedule creation time. Suppose a furnace goes offline 10 minutes after the start of the generation of a new schedule. Two things will happen:

a.  Schedule being built (unaware of the machine outage) may dispatch lots to the offline tool.

b. Machine outage will be handled only in the next schedule, 70 minutes after the machine went down.

2. Diffusion furnaces scheduled in isolation: Optimizing diffusion furnaces in isolation may cause other machines and areas to be neglected – resulting in suboptimal decisions. For example, since these clean tools feed other parts of the fab, there’s no guarantee that the necessary WIP will arrive at the furnaces to accommodate the optimized schedule having not taken clean capacity into account.

3. Assumption that transportation time of wafers is negligible compared to the processing time: despite the long processing times in furnaces, it’d be interesting to test transportation times in the model to confirm if it’s indeed irrelevant for scheduling or if it brings different decisions to the final schedule.

4. Loading and unloading time not addressed in the approach: Unlike processing times that are fixed, the loading and unloading times can still vary with the number of wafers.

Flexciton’s Way

Flexciton’s solution has been built to schedule any area of a fab through multi-objective optimization, handling multiple fab KPIs with their trade-offs and sending an optimized schedule to the fab every 5 minutes. Below, we outline the main features of how we tackle the main challenges of furnaces scheduling:

1. A fab-wide approach: our optimization engine schedules furnaces not in isolation but together with other machines across the fab. We utilise a holistic approach, looking ahead for bottlenecks across the entire factory and account for the existence of bottleneck tools when making scheduling decisions. For instance, a lower priority wafer may be dispatched before a high priority one if the former is going to a low-utilisation machine while the latter is going to a bottleneck in its next step.

2. Criticality of time constraints: whilst eliminating violations of timelinks, we account for the different criticalities they may present, be it because of the machines and recipes used or due to wafer priorities. This means that under a situation where one of two timelinks must be violated for reasons beyond our control, the less critical timelink will be violated.

3. Multi-objective optimization: We balance multiple KPIs simultaneously and handle their trade-offs through user-defined weights. For example, objectives such as “minimise timelink violations” and “minimise cycle time” can receive different weights depending on the desired behaviour in the fab. This directly impacts decisions such as “how long should a high priority wafer wait for a full batch?”.

4. New schedules every 5 minutes: Our technology is based on a hybrid approach that combines Mixed Integer Linear Programming (MILP) with heuristic and decomposition techniques, enabling the delivery of high-quality schedules to the fab every 5 minutes.

5. Change management: Adherence by operators and managers to a new scheduling system and its decisions is among the main post-implementation challenges. Because of that, our deployments follow a rigorous plan that helps foster a higher adoption of the technology. We also use detailed Gantt charts to aid the visualisation of schedules, which facilitates a solid understanding of decisions made which in turn enables higher adherence from operators.

As explored in this article, scheduling diffusion furnaces can be an extremely complex task. This is true even from a computational standpoint, leading many semiconductor fabs to rely on the judgement and experience of their operators at the cost of obtaining suboptimal and inconsistent schedules that take hours to generate. On the other hand, the usage of some fast-scheduling systems may mean leaving some constraints behind, ignoring different KPIs or not observing the fab in its entirety.

At Flexciton, we combine the best of both worlds and bring fast optimal decisions while fostering technology adoption at all hierarchies of the fab.

‍

In part one of this article, our case study illustrated how the use of advanced optimization can help balance the line by reducing queue times at bottleneck tools. By employing optimization, we were able to look into the future at the state tools further down the line and prevent problematic congestion.

In our second case study, we consider a more complex problem where a trade-off must be made between the cycle time of high priority lots and violating certain timelinks. The problem from the first case study was adjusted by increasing the priority of two of the lots (outlined in red in Figure 4).

Case Study 2: Balancing Cycle Time and Timelinks

In some cases, a fab may wish to favour the reduction of a high priority lot’s cycle time and violate a lower priority lot’s timelink. For example, violating certain timelinks may result in minor rework, such as re-cleaning, rather than resulting in scrappage. This rework takes time, but it could be a sacrifice a fab is willing to take with different priority lots.

Scenario 1: Optimizing to Meet Cycle Time Objectives (Benchmark)

When purely minimising cycle time without timelinks, the priority lots are brought forward as early as possible. If timelinks were present, then the final two batches of Figure 4 would violate their timelinks.

Scenario 2: Optimizing with Critical Timelinks

In this scenario, timelinks are introduced and are all considered to be critical. We eliminate their violations, but this comes with an 8% increase in cycle time. The priority lots highlighted in light blue in Figure 4 are already within a timelink at the start of the schedule. In order to meet the critical time constraints, the lots get delayed.

Figure 4 shows the difference between Scenarios 1 and 2.

Scenario 3: Considering Timelink Criticality and Cycle Time of High Priority Wafers

In this scenario, cycle time is a vital KPI for high priority wafers. Additionally, we consider two different levels of timelink criticality; critical (should not be violated) and non-critical (can be violated if necessary). Only the timelinks feeding the final step at the furnace tool are considered critical. The critical timelink violations are eliminated first, before optimizing non-critical timelinks and cycle time simultaneously. We configure the optimizer objective so that an hour of non-critical violations would be equivalent to an hour of cycle time for the lower priority lots.

Figure 5 shows the calculated schedule when these parameters are considered; all the critical timelinks between steps 5 and 6 are satisfied, however some non-critical timelinks are broken to reduce the cycle time of the high priority lots.

Figure 6 visualises the trade-offs between timelink violations and cycle time for the three scenarios discussed. It demonstrates a benefit of advanced optimization, where different schedules can be optimized depending on the goals of the fab – whether it's purely for cycle time or adhering to timelinks.  

Conclusion

Handling timelinks is crucial to managing the cost of running a fab, however it presents challenges. Our case studies demonstrate how Flexciton’s advanced scheduler can be used to plan several steps into the future when there are timelinks on bottleneck tools. Doing so helps spread WIP more evenly across tools and helps manage dynamic bottlenecks. By configuring priorities and the impact of missing timelinks, it can also flexibly trade-off violations of differing importance against each other and with other KPIs such as cycle time. This can be achieved without the need to rely on highly bespoke heuristics that are difficult to configure, require frequent maintenance and don’t necessarily guarantee the desired outcome.

‍

Timelinks (also known as time constraints, time lag constraints, time loops or close coupling constraints) are one of the most challenging aspects of a wafer fab to navigate and significantly increase the complexity of scheduling it. We take a dive into a case study that shows how optimization can be used to manage timelinks to alleviate pressure on bottleneck tools.

What are timelinks?

A timelink is a maximum amount of time that can elapse between two or more consecutive manufacturing process steps of a lot (a group of silicon wafers). Defining timelinks is necessary to mitigate the risk of oxidation and contamination of wafers while waiting between process steps. Violating these timelinks can lead to wafers being scrapped or undergoing a costly rework due to exposure to impurities – ramping up the production costs for a fab.

Managing timelink violations is therefore critical as its impact on cost and chip quality must be balanced with delivery speed (measured by cycle time) and other cost considerations such as leaving tools idle. When it comes to managing timelinks, most fabs look towards heuristics. However, the complexity of time constraints are difficult for a rules-based approach to navigate and require a more advanced scheduling solution. For example, deciding whether or not to dispatch a lot when the subsequent tool has a timelink requires the ability to look at the future state of tools further along the schedule. Alternatively, if you use optimization to generate a schedule then looking into the future becomes a lot more straightforward.

Using optimization to eliminate timelink violations

To demonstrate how optimization can help tackle the scheduling of timelinks on bottleneck tools we take a dive into a case study where scenarios are ran using Flexciton’s advanced scheduler. The case study demonstrates how timelinks at a bottleneck tool can be managed by looking several steps into the future and delaying earlier steps to more evenly balance the line.

For illustrative purposes, we consider a small problem with only 33 lots. Each lot has up to 6 remaining steps to be scheduled across 52 tools. There are time constraints of varying duration between all consecutive steps.

The sequence of process steps is defined by the product routes: A, B, C, D and ‘Other’. Routes A, B, and C all end on tool Z which is a diffusion furnace tool that runs batches of five lots at a time. All the timelinks around this tool are relatively tight at around two hours.

Case Study: Managing Timelinks at a Bottleneck Tool

We begin by running a production schedule through Flexciton’s scheduler without considering timelinks, where we prioritise minimising cycle time alone (Fig. 1).

When timelinks are included in the schedule, we first eliminate their violations before minimising the cycle time (Fig. 2).

Figure 2 shows the light green lots (the last batch on tool Z) are shifted right on the W toolset to avoid violating timelinks on the next step, and therefore incurring the cost of scrapping wafers or performing rework. This creates a period of idle time on the W toolset and delays the other lots on W. If toolset W was considered in isolation, this solution would be suboptimal. However, when taking into account both toolsets, this provides a far better outcome.  

With timelinks met, we also more evenly balance queue times across the line, as demonstrated in Figure 3, which shows total queue time at the final two toolsets used in Route A in order of process steps. Toolset Z still has the highest queue time due to it being the bottleneck, but the difference is substantially reduced. This balancing reduces the bottleneck effect on tool Z.

Conclusion

This case study illustrates how using advanced optimization can help balance the line and reduce the queue time at problematic bottleneck tools. By harnessing the power of optimization we are able to assess the state of tools further down the line – something that isn’t realistically possible with traditional heuristics. This ability to significantly reduce queueing time can go a long way to helping a fab manager to hit KPIs such as the reduction of cycle time whilst avoiding costs incurred from scrapping wafers. However, the problem of scheduling timelinks becomes even more complex when you begin to consider wafers of differing priorities.

Want to learn more? Take a dive into Part 2 of this article where we will be taking a look at how to solve this problem with the added complexity of priority wafers.

‍

A discipline of Machine Learning called Reinforcement Learning has received much attention recently as a novel way to design system controllers or to solve optimization problems. Today, Jannik Post – one of our optimization engineers – takes a look at the background of the methodology, before reviewing two recent publications which apply Reinforcement Learning to scheduling problems.

The exciting prospect of Reinforcement Learning

Traditionally, semiconductor fabs have relied on real-time dispatching systems to provide their operators with the dispatch decisions – with their ability to show the current state of the work in progress within seconds. These systems may follow rules based on heuristics or derive them from domain knowledge, which makes their design a lengthy process that requires deep knowledge of the fab processes. Maintenance of the contained logic also requires continuous attention from subject matter experts. As well as this, these systems have very limited awareness of the global effects of decisions at toolset level – therefore making them susceptible to providing suboptimal decisions.

More advanced approaches to wafer fab scheduling rely on optimization models, which can take many factors into account, e.g., the effect of dispatching decisions on bottleneck tools further downstream. These solutions will generally require a slightly longer computation time to achieve high-quality solutions.

Reinforcement Learning (RL) promises to avoid the downsides of both common dispatching systems and optimization approaches. So, how does it work? At the heart of RL there is an agent* which performs a task by taking decisions or controlling a system. The goal is to teach this agent to make close to optimal decisions by allowing it to explore different options and providing feedback on the quality of its decision. Good decisions are rewarded whilst suboptimal decisions are punished. Of course, this training will not be performed in a live environment, but rather by simulating thousands of scenarios that might occur to prepare the agent for any possible situation.

A common example of Reinforcement Learning is self-driving cars, but it can easily be seen how it could be productive when used in other environments, such as dispatching in a wafer fab. In theory, it could be utilised to dispatch wafers to tools in a way that optimizes certain KPIs – such as throughput.

Reinforcement Learning for Job Shop Scheduling** problems

Numerous recent publications have explored the use of RL for production control. However, the approaches are still in their early stages and applied to problems much less complex than semiconductor scheduling. Nevertheless, they demonstrate the potential to play a part in future solution strategies. Two approaches stood out to us when reviewing the literature:

“Learning to Dispatch for Job Shop Scheduling via Deep Reinforcement Learning” (2020)

This paper by Zhang et al. describes an approach to designing an agent that generalises its knowledge beyond what it has been trained to do, enabling it to handle unseen problem instances. This is achieved by initially conducting a large amount of diverse training scenarios. The model can flexibly handle instances of different sizes, e.g., with varying numbers of tools.

The agent is first trained on large numbers of scenarios and will thereby learn to exploit common patterns and perform well in instances not encountered before. After the training, the agent can be deployed to solve new instances. As training is conducted separately from solving an instance, the latter can be performed in less than a minute. The performance on benchmarking problems is compared against optimization models and simple dispatching heuristics. The Reinforcement Learning approach yields a makespan – the total duration of the schedule from start to finish – between 10-30% longer than when computed through optimization, but around 30% shorter than what simple heuristics achieve.

“A Reinforcement Learning Environment for Job Shop Scheduling” (2021)

This paper published by Tassel et al. sets out to design a reinforcement learning environment to optimize job shop scheduling (JSS) problems as an alternative to optimization models. The objective in this approach is to reduce periods in the schedule where tools are not in use, which is shown to correlate with a minimisation of makespan. The agent is designed as a dispatcher and is trained on a single scenario at a time by running a real world simulation over and over. As the goal is to generate an optimized solution for the instance, the best solution achieved during training will be saved. Training time and solution time are thus the same in this approach and are limited to 10 minutes to reflect production requirements. In this approach, there is no intention to generalise the behaviour of the agent to other instances.

The authors disclose a makespan of just 10-15% worse than the best known benchmarks for job shop scheduling, and just 6-7% longer than time-constrained optimization approaches.

Flexciton’s view

At Flexciton, we are excited about bringing cutting-edge optimized scheduling to wafer fabs worldwide. We are always exploring new ways that could help us improve the service we provide our customers so it’s exciting to see new emerging technologies which may help solve scheduling challenges in the semiconductor industry. The two publications reviewed in this article both present promising new approaches that yield measurable improvements over simple dispatching heuristics, but still fall short of optimization.

Both approaches can cope with disruption and stochasticity of the environment, such as machine downtimes. Another commonality is that both can readily be applied to problems of different sizes. In both cases the authors respected the requirement for frequent schedule updates (Tassel et al.) and quick decision support (Zhang et al.) and still achieved optimized solutions. It is conceivable that reinforcement learning has the capability to teach an agent to make smart decisions in the present that will improve the future fab state and reduce bottlenecks.

However, as the use of RL for JSS problems is still a novelty, it is not yet at the level of sophistication that the semiconductor industry would require. So far, the approaches can handle standard small problem scenarios but cannot handle flexible problems or batching decisions. Many constraints need to be obeyed in wafer fabs (e.g., timelinks and reticle availability) and it is not easily guaranteed that the agent will adhere to them. The objective set for the agent must be defined ahead of training, which means that any change made afterwards will require a repeat of training before new decisions can be obtained. This is less problematic for solving the instance proposed by Tassel et al., although their approach relies on a specifically modelled reward function which would not easily adapt to changing objectives.

Lastly, machine learning approaches can lead to situations where the decisions taken by the agent will be hidden in a black box. When the insights into the rationale behind decisions are limited, troubleshooting becomes difficult and trust into the solution is hard to establish.

Flexciton’s way

Using wafer fab scheduling to meet KPIs such as increased throughput and reduced cycle time is a challenge that requires a flexible, quick, and robust solution. We have developed advanced mathematical hybrid optimization technology that combines the capabilities of optimization models with the quickness of simple dispatching systems. When needed, the objective parameters and constraints can be adjusted without the need to rewrite or redesign extensive parts of the solution. It can therefore easily be adapted to optimize bottleneck toolsets, a whole fab or even multiple fabs.

Flexciton’s scheduling software produces an optimized schedule every five minutes and easily integrates with existing dispatching systems. The intuitive interface enables users to investigate decisions in a wider context, which helps during troubleshooting and increases trust in the dispatching decisions.

‍

References

[1] Zhang, Song, Cao, Zhang, Tan, Xu (2020). “Learning to Dispatch for Job Shop Scheduling via Deep Reinforcement Learning.”

[2] Tassel, Gebser, Schekotihin (2021). “A Reinforcement Learning Environment for Job-Shop Scheduling.”

[3] Five reasons why your wafer fab should be using hybrid optimization scheduling (Flexciton Blog)

Notes

* – We use the term ‘agent’ to describe a piece of software that will make decisions and/or take actions in its environment to achieve a given goal

** – The job shop is a common scheduling problem in which multiple jobs are processed on several machines. Each job consists of a sequence of tasks, which must be performed in a given order, and each task must be processed on a specific machine.

‍

The photolithography process is considered the most critical step in semiconductor wafer fabrication, where geometric shapes and patterns are reproduced onto a silicon wafer, ultimately creating the integrated circuits.  

What makes this process unique is the use of additional resources, called reticles. The reticle is a photomask used to expose ultraviolet radiation that generates a specific pattern into the wafer. When not in use, reticles are stored in dedicated storage with a fixed capacity, called a stocker. Although the problem of allocating reticles in a stocker is not a “core” one in wafer production scheduling, if not optimized, it might significantly impact overall production efficiency by causing bottlenecks.

This week, Daniel Cifuentes Daza, one of the Optimization Engineers here at Flexciton, explores this problem by reviewing a technical paper by Benzoni, A. et. al. – “Allocating reticles in an automated stocker for semiconductor manufacturing facility” – and contrasting their approach with the one we use when scheduling at Flexciton.

The reticle and stocker bottleneck:

A fab working with a wide variety of products may need several thousand reticles at any given time to fulfil their production requirements [2]. Not only must reticles be stored in a stocker, as explained above, but they often need to be transported in a container known as a pod in order to prevent contamination too. Therefore, the capacity and availability of stockers and pods within the fab makes deciding where each reticle should be stored at each step of the production schedule extremely complex – frequently causing bottlenecks [3].

To manage this process, fabs need to decide the best way to allocate reticles into pods, and then try to find an optimal assignment between pods and tools. However, there is another optimization problem at hand that complicates the process further; the position of reticles within the limited-capacity compartments of the stocker itself.

The time to retrieve a reticle from storage can be drastically different depending on its own location inside the stocker, thus leading to large inconsistencies in the so-called processing time of the stocker. As a result, the stocker can become a bottleneck by not dispensing reticles fast enough to meet wafer demand. Therefore, the reticle allocation problem also consists of choosing which reticles are to be stored in the low-capacity fast-retrieval compartment (“retpod”) vs the high-capacity slow-retrieval compartment.

In order to explore what might be the best way to address this problem, we have reviewed a tech paper published by IEEE for the WSC conference in 2020. The authors of the paper address the allocation issue using the famous knapsack problem. This approach will be evaluated in the next section of this article – with the pros and cons being discussed – before discussing how the proposed solution compares to how we model photolithography tools here at Flexciton.

The reviewed paper approach:

In “Allocating reticles in an automated stocker for semiconductor manufacturing facility” by Benzoni, A. et. al. (2020) [5] the stockers examined by the authors have two compartments – one where reticles are stored using pods; retpod, and another where they do not use pods. The main objective is to allocate reticles into the retpod compartment, as this has faster retrieval times.

Additionally, the authors consider:

(1) the reticles, as the main resource of the problem,

(2) the steps where wafers will need to use the reticles in the near future, and

(3) the capacitated storage for reticles; the compartments of the stockers. Additionally, they see each reticle as having a profit value; the number of wafers processed in the batch. With this initial information, the problem can be modelled as the well-known knapsack problem.

Sheveleva, A. et. al (2021) [4] defines the knapsack problem as the following:

“There are k items with weight nk and value ck, and a knapsack with a capacity N. The problem is to fill the knapsack with items with the maximum total value, respecting the knapsack’s capacity limit”

In this case, each k item is a reticle, where its corresponding weight is always 1, and its value ck refers to the profit value. The knapsack is the retpod compartment of the stocker.

The knapsack problem is an NP-hard combinatorial problem that has been studied for many years within computer science, operations research, and other sciences. Therefore, due to its complexity, the authors decided to use a well-known heuristic. Here, the approach is to rank each reticle according to a specified objective value ratio and then fill the knapsack with the first N elements fulfilling its capacity.

The authors benchmarked three different objective functions for this heuristic as follows:

1. Reticles are ranked according to their priority and their probability of being used soon. Reticles required by wafers in the short term have higher priority than those needed in wafers for later steps.
2. Similar to (1), but the ranking also considers the total number of wafers that each reticle is being used by.
3. Ranks reticles by the priority and probability of the reticle to be used, but only considers the most immediate reticles being used by at least one wafer.  

The three approaches reported an increase in the utilisation of the reticles from around 8% to 20%. This implementation also led to a reduction of processing times for the stockers of 1 hour. Strategies 1 and 2 showed the lowest error percentage, which is expected as Strategy 3 does not consider future steps where reticles are used.

Flexciton’s view:

Using the knapsack approach to solve this problem certainly has some positive points. Firstly, using a heuristic method is easy to implement and does not require much computational time, which also makes it scalable to industrial-sized problems. Secondly, it is trivial to work out why certain allocation decisions are taken, making it highly understandable. Lastly, the approach is flexible because the user can modify the objective function of the heuristic depending on the fab’s goals.

However, the issue of reticle allocation is just a small piece of the complex wafer manufacturing process. Since this approach is modelled as a standalone problem, it is creating feasible solutions for the reticle stocker alone without considering the state of the rest of the fab. This will likely lead to inconsistencies as the wafer schedule is intrinsically linked to the reticle allocation.

In addition, the approach described in the paper models a simplification of the photolithography area. There is relevant information missing, such as the availability of pods in the fab and their possible allocation to machines, transfer times, and load and unload times. The use of this information would give robustness to the approach.

Flexciton’s way:

At Flexciton, we consider that the best way to tackle the reticle allocation problem is to proactively generate not only feasible solutions, but optimized production schedules. In order to do this, we take into account all the scheduling constraints for reticles available within our optimization engine – using information such as:

• Availability and capacity of pods for reticles
• The scheduling of pod usage – taking into account the number of pod ports in the machine
• Storage of pods into stockers, considering its capacity
• Transition of pods between stocker and machine
• Load, and unload times for reticles into pods and pods into machines
• Variable stocker unload time on account of where the reticles are located
• The reticles and pods already loaded into the machines; where they are in the fab at the time of scheduling

‍

The benefit of considering a multitude of information like this all in one optimization model means that we can provide a consistent and robust production schedule that takes into account all the constraints of reticles, pods and stockers. Additionally, our scheduler allows the user to configure their specific business objectives into the optimization process in order to meet their fab’s KPIs – the algorithm is then calculated and an optimized schedule is returned in a matter of minutes. All of this means that our technology is able to return a reliable, scalable and flexible solution that is tailored to our client’s needs – whilst optimizing the photolithography area in its entirety.
‍

References:

[1] Y. T. Lin, C. C. Hsu and S. Tseng, "A Semiconductor Photolithography Overlay Analysis System Using Image Processing Approach," Ninth IEEE International Symposium on Multimedia Workshops (ISMW 2007), 2007, pp. 63-69, doi: 10.1109/ISM.Workshops.2007.16.

[2] S. L. M. de Diaz, J. W. Fowler, M. E. Pfund, G. T. Mackulak and M. Hickie, "Evaluating the impacts of reticle requirements in semiconductor wafer fabrication," in IEEE Transactions on Semiconductor Manufacturing, vol. 18, no. 4, pp. 622-632, Nov. 2005, doi: 10.1109/TSM.2005.858502.

[3] You-Jin Park and Ha-Ran Hwang, "A rule-based simulation approach to scheduling problem in semiconductor photolithography process," 2013 8th International Conference on Intelligent Systems: Theories and Applications (SITA), 2013, pp. 1-4, doi: 10.1109/SITA.2013.6560788.

[4] A. M. Sheveleva and S. A. Belyaev, "Development of the Software for Solving the Knapsack Problem by Solving the Traveling Salesman Problem," 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus), 2021, pp. 652-656, doi: 10.1109/ElConRus51938.2021.9396448.

[5] A. Benzoni, C. Yugma, P. Bect and A. Planchais, "Allocating Reticles in an Automated Stocker for Semiconductor Manufacturing Facility," 2020 Winter Simulation Conference (WSC), 2020, pp. 1711-1717, doi: 10.1109/WSC48552.2020.9383933.

‍

In order to maintain high margins, the cost of manufacturing semiconductors needs to continually diminish. Previously, this had been achieved by increasing the wafer size and shrinking the size of the chips whilst increasing the density of transistors. However, as the effects of these tactics begin to culminate, the future of maximising a wafer fab’s capacity lies within optimizing operational processes.

The process of manufacturing a semiconductor chip is exceedingly complex, often requiring thousands of unique steps. Reducing cycle times and increasing throughput in such an intricate production process calls for very high production efficiency. Fabs usually approach this in one of two ways; the heuristic approach, which is fast but not optimal and the mathematical approach, which is optimal but time-consuming. In order to attain optimal results that are able to keep up with changes on the factory floor, however, fabs need to switch to advanced production scheduling.

The hybrid approach

At Flexciton, we are pioneering a new model that combines the two different methods with a hybrid technique. With this ground-breaking new model, which we call advanced mathematical hybrid optimization technology, optimal results are delivered in a matter of minutes.

Here are 5 reasons your fab will benefit from switching to an advanced scheduling solution:

It's quick

The most complex of scheduling problems can be solved in less than 5 minutes, delivering near-optimal results. This makes the hybrid technique perfect for the dynamic fab environment, since updates can keep up to speed with changes on the factory floor.

Accommodates all constraints

Hybrid optimization is able to realise fully accurate schedules by accommodating all constraints. This ensures a true representation of all activity in the fab, as well as its limitations.

Optimal schedules

By employing mixed-integer linear programming (MILP), Flexciton’s hybrid method guarantees high-quality solutions. Thanks to performance enhancing decomposition, final solutions are very near to the global-optimal.

Low maintenance

With MILP being the core of the solution, high performance can more easily be maintained with very little upkeep. With changes in objectives and recipes constantly taking place, not every consecutive shift in a fab is alike. Despite this, Flexciton’s solution can take into account all aims and constraints and consistently calculate an optimized schedule.

It's adaptable and easy to alter

When needed, constraints and parameters can be altered without the need to rewrite or redesign extensive amounts of new code. Not only this, but hybrid-optimization scheduling can also be rolled out into the entirety of a fab (global scheduling) as well as multiple fabs, even when they have differing production characteristics.

Hybrid optimization could be the answer to your fab’s scheduling problems. Download our white paper to find out more.

‍

Introduction

Photolithography processes are central to producing computer chips and semiconductor devices.  However, they are typically considered to be bottlenecks due to their reliance on a critical secondary asset; reticles. Reticles are limited in number and yet are a critical piece of the coat-expose-develop loop. What is more, reticles are delicate in nature; they are enclosed in purpose-built cases for their transport in order to keep the potential of damage or distortion to a minimum. 

As such, a fundamental tradeoff arises when operating photolithography toolsets: moving a reticle to the machine it is needed most (to carry out high-priority tasks) clashes with the requirement to be conservative with its transport. In theory, there are several compromises that the operator can make to reduce reticle movements - waiting a bit longer to ensure more wafers arrive to a machine and a larger batch can be processed with a single move is one example. However, in practice, identifying these strategic actions and balancing between the competing goals is highly complex. Flexciton can provide a solution to this issue by leveraging the power and flexibility of optimisation.

In this article, we show how the Flexciton’s scheduling engine can balance between minimising cycle times and reticle moves. Through a series of example case studies, we delve into the scheduling trade-offs that arise in the day-to-day operation of a semiconductor fab and how Flexciton’s solution can assist in uncovering schedules that optimally balance across competing goals. 

Trade-off frontier with the Flexciton scheduler​

The Flexciton scheduler can accommodate a range of user-defined objectives. The fab operator is typically interested in minimising KPIs such as cycle times but may also want to include other considerations, such as penalisation of labour-intensive decisions e.g. number of batches built. In this vein, we have recently introduced a new component; the number of reticle moves carried out. As shown in Figure 1, the user is able to define a penalty factor for reticle moves; the higher the value, the harder the engine will try to avoid moving reticles. 

​

In the example case study we have 6 machines and 48 wafers to be scheduled, with a total of 4 reticles.

Reticle 10001 is required for all lots schedulable on machines 01, 02 and 03. Deciding on how to move this in-demand reticle across the 3 machines will impact both the number of moves as well as cycle times, particularly as there are some high-priority wafers waiting to be dispatched. Reticle 10001 is originally loaded in machine 01.
‍
The other three reticles, 20001, 20002 and 20003 are initially loaded on machines 04, 05 and 06 and can be used by all three toolsets interchangeably. However, different machines are better suited to different reticles; for example in our case study, the same process is completed faster if using reticle 20001 on machine 06. Note that all machines have a maximum batch size of 4 wafers.

Case Study 1: Focusing solely on cycle times​

We start off by not penalising the number of reticle moves and solely minimising the total priority-weighted cycle times (TWCT) across all wafers. The optimal schedule produced by Flexciton’s engine is shown in Figure 2. 

There are a total of 7 reticle moves, noted with red arrows in the figure below. 4 moves pertain to reticle 10001 which is moved from its initial location 01 to 03 and then 02 to carry out some high-priority wafers (as evidenced by the circled 1/2/3 next to the job names). The reticle is then moved out again to machine 01 and finally to machine 03 to carry out some lower priority jobs. Looking at machines 04, 05 and 06, the engine decides to immediately swap the reticles between the machines, to ensure that each lot is fed to its most suitable (in terms of processing times and capability) machine. The TWCT of all 48 wafer steps (where priority weights are user-defined and in this case range from 1 for highest-priority to 0.1 for lowest-priority wafers) is 16.79 hours.

Case Study 2: Moderate penalisation of reticle moves

In this second study, we have penalised reticle moves; the ratio for balancing TWCT and reticle moves has been set to 100:75 i.e. we will choose to avoid a reticle move only if its avoidance translates to an increase of TWCT of 0.75 hours or less. This is quite relaxed, but is aimed at avoiding reticle moves with little benefit, since the risk of potential damage is deemed higher. The optimal schedule obtained is shown in Figure 3.

In this study, there are a total of 5 reticle moves, noted with red arrows in the figure below. 3 moves pertain to reticle 10001 and its journey across machines 01, 02 and 03. The main difference to the previous scheduling pattern is that now we do not move the reticle back to machine 03 to carry out the very last batch of low priority wafers. Instead, we choose to wait for their arrival and carry them right after the high-priority batch finishes a bit after 19:00. This way we avoid that final reticle move, while also incurring a delay in the high-priority wafers scheduled on machine 02 which now have to be moved from 19:15 (in study 1) to 19:30.

Looking at machines 04, 05 and 06, the engine decides to immediately swap only two the three reticles this time, and leave reticle 20001 on its initial machine. Although that initial setup is not ideal in terms of processing times it does prevent the reticle move deemed to be “lower value”. The TWCT of all 48 wafer steps is 18.73 hours. 

‍

​​Case Study 3: Only necessary reticle moves

In this study we look at the extreme case of using a very high penalty on reticle moves, hence allowing only absolutely necessary reticle moves. In particular, we have opted to use a TWCT to reticle move cost ratio of 1:10. In such cases, the operator is willing to accept sub-optimal job-machine allocation decisions, as well as delayed scheduling of high-priority wafers, for the purpose of keeping reticle movement to the absolute minimum. The optimal schedule obtained is shown in Figure 4.

In this study, the total number of reticle moves has come down to just 2 moves, noted with red arrows in the figure below. Both moves pertain to reticle 10001 and its journey across machines 01, 02 and 03 to ensure all wafers are completed. In the case of machines 04, 05 and 06, we are still able to carry out all tasks, albeit with longer processing times, as evidenced in the much later finishing times of the machines. The TWCT of all 48 wafer steps is 23.20 hours. 

​

​Exploring the trade-off frontier

Plotting the aforementioned runs (and also some more data points), we obtain Figure 5, which clearly illustrates the trade-off at play here. As we traverse the penalty factor from a low to a high value, the number of reticle moves drops and the cycle times increase. As expected, these relationships are monotonic but not smooth, since they depend on discrete events. Note also that both curves are bounded both from above and below, corresponding to the absolute minimum number of reticle moves required (in this case 2) and the absolute maximum number of reticle moves that is optimal (in this case 7).

By running a few scenarios with different parameters, the Flexciton engine opens up the possibility to explore the tradeoff frontier in detail, enabling operators to quantify how KPIs would change with a more relaxed or constrained attitude towards reticle movements.

​

​Performance in real-world applications​

In practice, enabling the Flexciton scheduling engine to consider reticle moves is a computationally challenging task, involving novel development in the model’s MILP formulations and heuristics. Nevertheless, this feature has been accommodated with no deterioration to performance and schedule quality. The Flexciton engine is capable of scheduling thousands of wafers across hundreds of machines in a few minutes while also controlling for the operator’s tolerance to reticle movement. 

Indicatively, we showcase results obtained from scheduling a real-world fab plant. At the time of the study, the plant had a total of 3,478 wafers to be scheduled on 209 toolsets (with a total of 358 load ports). We computed two schedules: one with low and one with high penalisation of reticle moves. These scheduling runs were computed in roughly the same time: respectively, confirming that despite the added complexity, this feature can scale well and provide a schedule in a few minutes.

Focusing on the reticle machines, the results obtained suggested that reticle movements could be reduced by around 26% while leading to an increase in total cycle times of around 2%. Note that these results are priority-weighted, with further analysis revealing that high-priority wafers are not substantially impacted; the optimiser is able to identify “low-value” reticle movements relating to e.g. early processing of a low priority wafer and either avoid that movement by using an alternative recipe, or deferring that movement to later when a low-priority wafer can be combined with a high-priority wafer in a batch.​

Conclusions

Reticle scheduling is a very important consideration in the scheduling of advanced semiconductor fabrication plants. This resource, already highly constrained, comes with a critical consideration in practice: frequent movements and manual handling of the delicate reticles increase the risk of damage or distortion during transport. As such, the number of times a reticle is moved to a new machine must be managed conservatively. This inadvertently clashes with the operator’s fundamental objective of reducing cycle times. 

Flexciton has extended the capabilities of our Mixed Integer Linear Programming (MILP) scheduling engine to natively accommodate the modelling and penalisation of reticle movements. This allows the user to define their own risk profile, so as to limit reticle movements solely to cases deemed of high value. In addition, the engine opens up the possibility to explore this tradeoff frontier in detail, enabling operators to quantify how their plant’s performance may change with a more relaxed or constrained attitude towards reticle movements.

Authors

Ioannis Konstantelos is a Principal Optimisation Engineer at Flexciton. He holds a PhD from Imperial College London and has published over 50 conference and journal papers on optimization and artificial intelligence methods. Ioannis joined Flexciton over 3 years ago and is involved in the development of Flexciton’s scheduling engine.
‍
Charles Thomas is a Test Analyst with a background in Mechanical Engineering and a Masters degree from the University of Southampton. He has been at Flexciton for 2 years and leads the benchmarking and testing of the application with a particular focus on scheduling engine performance.

‍

Reviewing technology literature is a common practice when developing a new approach to solving an existing problem. James Adamson, a Senior Optimization Engineer at Flexciton, has recently reviewed several technical papers on photolithography scheduling, one of which he found particularly interesting.

The reticle challenge

Photolithography is the cutting edge of semiconductor manufacturing and as a result, requires the most complex and expensive equipment to run and maintain. Reticles (also known as photomasks), must be prepared and loaded into litho tools before the wafers can process. These fragile masks are extremely expensive (in the region of $100k or more [Weber 2006]), making them a scarce resource.

Wafers require specific reticles for their individual process step. Therefore fab operators need to ensure that the correct reticles are at the correct tools on time in order to keep production KPIs such as wafer throughput and cycle time optimal. While reticles can be moved between tools, this takes time and considering how fragile these masks are, the movement needs to be minimised as much as possible. If wafer scheduling wasn’t already difficult enough, we now have to wrestle with reticle scheduling too.

The approach review

The paper “A Practical Two-Phase Approach to Scheduling of Photolithography Production” by Andy Ham and Myesonig Cho was published in 2015. The authors present an approach that is based on the observation that most semiconductor manufacturing companies are still using real-time dispatching (RTD) systems to make last-second dispatching decisions in the fab. RTD has the advantage of being familiar and relatively understandable whilst also being fast computationally. In contrast, some optimization-based approaches, particularly for photolithography, can struggle to scale up to industrial-scale problems. The authors' approach exploit the idea that an exact schedule for the next several hours is not necessarily needed and that RTD will ultimately be responsible for the final dispatch decision.

They propose a two-stage approach that integrates a simple heuristic (designed to mimic a fab’s RTD system) with mixed-integer programming (MIP):

1. Stage 1 - Upper-stage (Assignment): MIP is used to guide the high-level lot-to-machine and reticle-to-machine assignments. Explicit timing of lots and reticles is ignored.
2. Stage 2 - Lower-stage (Sequencing & Timing): The finer-detailed sequencing (and therefore timing) of lots according to those assignments is provided by the heuristic.
   ‍
   

The two stages are then tied together in an iterative fashion. A set number of lots are scheduled in each iteration of the two stages. The algorithm then repeats from Stage 1 with new additional lots and keeps iterating until all lots are scheduled.

Stage 1

A MIP approach is proposed to solve the assignment problem in the first stage. Two primary decision variables are used:

• A variable that allocates a reticle to a machine
• A variable that allocates a reticle to a lot on a machine
  
  ‍

The model does not account for the explicit timing of lots on their allocated machines. Therefore it cannot prescribe a sequence of lots or reticles on the machines. It just indicates that they will be scheduled on this machine at some point. The model requires that all lots are assigned to a machine and a reticle. Finally, the model measures the completion time of each machine as a function of the processing time of all lots allocated to the machine, rather than explicitly deciding the order of each lot on the machine.

Multiple objectives are used to achieve the trade-off between reticle movements, cycle times, and machine load balancing:

1. Minimise completion time; the end of the last lot’s process
2. Minimise the difference between the earliest and latest machine completion times. This aims to achieve load balancing across machines
3. Minimise total machine underachievement, where underachievement is the gap between the total processing time on a machine and the average total processing time over all machines.
   

Stage 2

Sequencing decisions are handled by an RTD system, where the manufacturer’s custom business rules can be applied; however, the lot-reticle-machine assignment decisions are fixed. This reduces the scope of the decision-making that RTD must make. The authors highlight the benefit of explainability with this two-stage approach. When questioning assignment decisions, the assignment model should be explored, whereas when questioning sequencing decisions, RTD should be investigated.

This practical approach was shown to solve reasonable problem sizes (500 lots, up to 800 reticles, 30 machines) in 2-4 minutes. They managed to reduce cycle times by 3%, on average, and, particularly interestingly, reticle movements by up to 40% when compared to standalone RTD.

Flexciton’s View

Although the model does have some shortcomings, as outlined above, the practicality of the approach makes it a strong candidate for production-size scheduling as very few studies have been able to effectively handle industrial problem sizes for photolithography tools. The reduction in reticle movements achieved, in particular, cannot be ignored.

However, with the notion of time largely ignored in the assignment model, the approach outlined is certainly simplistic.

There are a number of factors not considered in the model, including:

• The effect of batching decisions on processing times of the lots. For example, if 5 lots could be assigned to machine A as a single batch, and 5 different lots were assigned to machine B in five unique batches, then the assignment model would consider these cases as having equal total processing times (which is unlikely to be true).
• The sequencing (and therefore timing) of reticles on machines.
• The batching of reticles into “pods”; a container of reticles for easy and safe transportation around the fab and within lithography machines such that they can’t always be separated
• How high priority wafers are handled in relation to lot-reticle-machine assignments
• Accounting for periods of downtime or machine unavailability, which would contradict with the aim of achieving machine load balancing.
  ‍
  

Flexciton’s Way

At Flexciton, we schedule a variety of photolithography tools as part of our optimization engine. Our hybrid optimization-based solution strategy is therefore capable of handling all the intricacies of a wafer fab simultaneously, such as the issues described in the previous section;

• modelling photolithography tools that require pods,
• priority lots,
• batching tools with variable processing times based on the batch, and
• machine unavailability.
  
  ‍

Not only do we model these complexities, but we also succeed in achieving high-quality schedules in little computation time.

The user is given the option of controlling various relative priorities of the lots, in addition to deciding the relative importance of KPIs such as reticle movements vs lots’ cycle time. The flexibility of an optimization approach that considers all of the advanced photolithography constraints combined with a self-tuning model that has limited tuning parameters is what makes our engine highly attractive as a semiconductor scheduler.

1) Weber, C.M; Berglund, C.N.; Gabella, P. (13 November 2006). "Mask Cost and Profitability in Photomask Manufacturing: An Empirical Analysis". IEEE Transactions on Semiconductor Manufacturing. 19 (4). doi: 10.1109/TSM.2006.883577

‍

Time constraints (also known as timelinks) between consecutive process steps are designed to eliminate queueing time at subsequent steps. In a highly complex wafer fabrication environment, even the most advanced fabs struggle with scheduling time constraints. While our engineering team works on applying Flexciton technology to solve the timelinks problem, Begun Efeoglu Sanli, one of our Optimization Engineers, reviews a recently published technical paper on this particular subject.

A fab manager perspective: time constraints and a fab performance

A silicon wafer undergoes a fabrication process by entering multiple production steps, where each step is performed by different, highly sophisticated tools. Optimizing the transition and waiting time of the lots has a huge impact not only on a fab production performance but also its profitability. As an example, by introducing time constraints at the wet etch and furnace process steps, we prevent the likelihood of oxidation and contamination. Failing to do so, risks contact failures, low and unstable yields, the consequence of which is either rework or the wafers must be scrapped. Such problems are difficult to discover during wafer processing, and to run special monitoring lots would be a considerable effort.

Yield optimization has long been considered to be one of the key goals, yet difficult to achieve in semiconductor wafer fab operations. As the semiconductor manufacturing industry becomes more competitive, effective yield management is a determining factor to deal with increasing cost pressures. Time links between consecutive process steps are one of the most difficult constraints to schedule, with a significant impact on yield management.

Some factories avoid the problem by dedicating tools to each process group that requires a previous cleaning or etch step. This strategy's obvious disadvantage is the higher demand from wet tools, which leads to higher investment, more cleanroom space, and ultimately to lower capital efficiency. The tradeoff between increasing throughput and a higher likelihood of violating lots’ time constraints is an everyday battle for fab managers trying to meet yield targets.

Technology perspective: scheduling time constraints

An example of time constraints for a single lot is illustrated in Figure 1 below. It shows a time link system between four consecutive process steps. In this example, we can see that the lot has time links constraining Step 2 to Step 4 as well as from Step 3 to Step 4, with overlapping time lag phases (also known as a nested time link constraint). This means that after completing process Step 3, the lot begins a new time lag phase (Time Link 3) whilst already transitioning through an existing time lag (Time Link 2) started upon completion of Step 2. As you might expect, the need to simultaneously look ahead and consider future decisions whilst also being constrained by past decisions is not trivial to model well in a heuristic or as real-time dispatch rules.

Time constraints are already difficult to navigate, but nesting them adds yet another layer of complexity for heuristics to wrestle with. In the example below, if the final step cannot be brought forward, scheduling Step 3 too close to Step 2 may make it impossible to meet the "Time Link 3".  It is because the time between Steps 3 and 4 is now greater than the maximum allowed. This would not be a problem if the time constraints were not nested and we only have to schedule according to the "Time Link 2".

The technical paper review

Surprisingly, although time constraints are an important topic, they have not been widely discussed in the tech literature so far. That said, an interesting paper on this topic was presented at the Winter Simulation Conference 2012 in December, by A. Klemmt and L. Mönch, “Scheduling Jobs with Time Constraints between Consecutive Process Steps in Semiconductor Manufacturing” of Infineon Technologies and University of Hagen.

The authors propose mixed-integer programming (MIP) model formulation and share some preliminary experimentation. Unfortunately, even state of the art MIP solvers can only solve problem instances up to 15 jobs and 15 machines to optimality in a reasonable amount of time.

Consequently, the authors develop two alternative approaches:

1. A heuristic that focuses on creating a feasible schedule where all time constraints are respected
2. A mathematical approach based on their MIP model that extends the heuristic in (1). The idea is to break the overall problem down into many smaller optimization problems that are easily solvable individually. These problems aim to ensure all lots are delivered on time and no time constraints are violated.

This novel decomposition approach allows for solving considerable problem instances including more than 100 machines, more than 20 steps, nested time constraints, and a large number of jobs.

Flexciton’s View

The paper referenced above highlights that both approaches can provide good feasible schedules quickly and as expected, the MIP-based heuristic outperforms the simple heuristic. Nevertheless, as with most heuristic approaches, there are some important tuning parameters that might affect the schedule quality. In the paper, the lots are sorted with respect to their due dates to build subproblems. This approach could perhaps be reevaluated if cycle time is the most important KPI for a fab or if some jobs are of higher priority. Similarly, if time constraint violations are allowed to some degree, then one could relax the importance of this in the heuristic.

The most important consequence, which is also mentioned, is that the cycle time of time-constrained processes correlates highly with utilization of upstream (start of the time link) processes. This eliminates waiting time in front of the upstream tool and enables higher utilization. However, if downstream tools are bottlenecks, then WIP may have to be withheld such that time constraints are not violated by stagnating in front of busy tools.

Another tradeoff to be considered is how low-priority time-constrained steps are scheduled among high-priority non-time-constrained steps. For example, is it worth risking a time constraint violation for the sake of rushing an urgent lot through the toolset? This needs to be quantified and considered by the fab manager. Therefore, all these tradeoffs should be taken into account in order to provide the best schedule.

The Flexciton Way

At Flexciton, we include time constraints as part of our optimization engine that attempts to eliminate all violations of time links as the highest priority. Only as a last resort are time constraints relaxed if it is not possible to provide an otherwise “feasible” schedule. This could occur if the time windows provided are unrealistically short considering all operational constraints, especially tool capacities.

As mentioned in the previous publications, the Flexciton optimization engine is a multi-objective solution that can balance various KPIs according to user-chosen weights, one of which controls the degree to which violations of time constraints are penalized. The main advantage of this approach is that, with all the other competing objectives, our solution can balance throughput, cycle time and priority-weighted time constraint violations simultaneously.

‍

In the constant pursuit of improved efficiency in semiconductor wafer fabs, the reliability of equipment is essential. The tools used in a fabrication process are extremely sophisticated; requiring an extensive preventive maintenance regime to ensure reliable production. A big challenge faced by fab managers is getting in place optimal scheduling of preventative maintenance whilst still meeting their production KPIs. Simply because such scheduling is extremely complex, involving many trade-offs as well as being time-consuming. What’s more, the exact impact on productive output is difficult to quantify.

Typically, to handle this complex problem, a fab may develop statistical models that try to predict unexpected tool downs. Such preventive maintenance – based on a pre-determined frequency – can help to minimise unexpected disruptions.

However, determining optimal maintenance frequencies is not an easy task, requiring answers to numerous questions and trade-offs that impact the eventual ability of a fab to meet its KPIs. Such questions include:

• How do different frequencies in maintenance impact the fab’s ability to meet on-time-delivery?
• How does one decide which tools to take offline at a single time, to minimise the disruption to the fab overall?
• Is it possible to forecast whether taking one arrangement of tools offline will lead to a 4% drop in throughput for the day? Whereas an alternative arrangement may only yield a 2% drop?
• What is the impact of taking those same tools down at 7 am tomorrow as opposed to 5 pm two days from now?
• What if – due to personnel requirements - you need all maintenance to take place at the same time? Alternatively, can they be staggered in 10-minute intervals for the same reason?‍
  

Fulfilling your KPIs

But what happens if it was possible to use your KPIs as basis to optimize maintenance scheduling? Instead of using a simple rule-based predictive model, such scheduling weighs constraints and finds the optimal schedule that will enable you to meet your KPIs.

Flexciton's scheduling technology addresses all such questions by finding the optimal schedule for your fab in any variety of forecasted conditions.

A ‘what-if’ scenario capability allows fab managers to effortlessly trial new preventative maintenance plans based on a variety of trade-offs or constraints. In addition, rather than dictate the time that tools must be taken offline, our optimizer will ensure all KPIs are achieved as best as possible, given the constraints.

By doing so, it prescribes the optimal maintenance schedule for the factory. All the fab manager has to do is to decide on suitable windows of time for each of the tools to be taken down.

Smart scheduling technology in practice

Let’s see what happens in three scenarios where we apply Flexciton’s maintenance scheduling capabilities with varying degrees of scheduling complexity. The scenarios are structured as follows:

1. The status quo — In the first case, we optimize a production schedule, with fixed maintenance timings prescribed to all start at the same time for a given toolset. Production is scheduled using a heuristic-based dispatch system emulating that found in many fabs.
2. Optimizing production around fixed maintenance — We use Flexciton’s advanced optimizer to perform the production scheduling.
3. Simultaneously optimizing production and maintenance — Finally, we allow the Flexciton advanced optimizer to schedule both production and maintenance timings. The window to flex the maintenance timings is chosen to be a 90-minute addition to the original timing provided in Case 1.
   ‍
   

The Gantt chart below (Figure 1) shows a snapshot of 300 lots scheduled in small toolsets over the course of twelve hours. Each lot can only go to a certain number of tools within that toolset, where the toolset is identifiable by the tool’s prefix. Each lot is assigned a priority. We optimize for the total cycle time of the lots, weighted by their priority. The maintenance periods (shown in striped orange) are of varying duration and are randomly assigned to tools to take place at a specific fixed time somewhere in the twelve-hour schedule.

In Case 1, we compare the logic of scheduling, given these fixed maintenance timings, with a heuristic dispatcher against Flexciton.

Here we can see that on the ‘XZMW/097’ tool, the dispatch system struggled to ‘look ahead’ and dispatch effectively, when given obstacles such as upcoming downtime just after 02:00. It would be better to, in the meantime, dispatch a short processing lot. An even more ideal schedule can flexibly move downtime around to maintain consistent, predictable throughput across the schedule.

So, what if the scheduler is allowed to prescribe the timings that it finds optimal? The following Gantt chart is from Case 3, where the optimizer is free to plan the maintenance at any time within a 90-minute window.
‍

The benefits of a flexible maintenance approach

To get a quick understanding of the results achieved using the three scenarios, we use queue time as our KPI. In the table below, you can see that the flexible maintenance approach greatly outperforms a simple dispatch heuristic. Obviously, queue time is a one-dimensional constraint of which there are infinitely more in a fab process that need to be considered. It is here that our maintenance optimization solution offers fabs unique capabilities: weighing all possible constraints to ensure KPIs are met to the fullest.
‍

So instead of letting predictive maintenance schedules drive production, why not let the driver of maintenance planning be your fab's top-line production KPIs? The Flexciton optimizer allows easy scenario testing and exploration in order to effectively quantify the impact that maintenance has on the production schedule.

Flexciton’s solution enables fab managers to consider multidimensional trade-offs simultaneously. The alternative to such informed decision-making is that fabs schedule their maintenance ‘blind’. They will ultimately pay the price through unpredictable cycle time, unsatisfactory throughput and unnecessary tool downtime. By switching over to smart scheduling, it is much easier to get an accurate prediction of the impact that modifying a downtime schedule will have in terms of meeting top-level KPIs. Learn more about smart scheduling by downloading our white paper, "Superior Scheduling: hybrid approach boosts margin"

‍

Preventive maintenance is a common practice in semiconductor wafer fabrication and essential for overall equipment availability and reliability. A typical approach is to plan maintenance activities ahead of time using simple rules-based models, where the maintenance is run on a particular day, at a particular time. The consequence of such approach, however, is optimising maintenance timing at the expense of production KPIs such as cycle time and throughput. What if we consider it the other way around and treat these KPIs as priority in the objective?

The maintenance scheduling problem

The complexity of semiconductor wafer fabrication entails a huge number of decisions and trade-offs a fab manager has to deal with each day. Preventive maintenance is one of them. The equipment used in the fabrication process is extremely capital intensive; therefore, it is critical that tools are utilised effectively and maintained on a regular basis to avoid any failures. Any servicing requires stopping the tools and suspending it from the manufacturing processes for a given period of time. With a recent shortage of chips, for the automotive industry, in particular, a fab manager faces a significant challenge - how to schedule preventive maintenance operations whilst ensuring maximum OTD and high throughput?

Maintenance scheduling is an established topic of research, with many authors showcasing various ways of solving this scheduling problem using simulation and optimisation techniques. An interesting tech paper on this topic was presented at the Winter Simulation Conference 2020 in December, by A. Moritz et al. “Maintenance with production planning constraints in semiconductor manufacturing” of Mines Saint-Étienne and STMicroelectronics. [the paper]

In this article, Ioannis Konstantelos, our Optimization Technology Lead, reviews the paper and explains Flexciton's approach to this complex topic.  

Technology perspective

The authors focus on identifying the best possible period of time (e.g. a day), across a large time range, in which to carry out maintenance tasks, while “respecting production deadlines and the capacity constraints on tools”. Two mathematical models are presented; in model 1, the maintenance is seen as a task that must be performed in a single period of time, e.g. one day (24h), while model 2 allows maintenance to be distributed across two consecutive periods, e.g. two days (48h).

Both models treat production schedules as fixed i.e. the lot to tools  assignment and timings for production purposes have been decided a priori. As such, the proposed formulation is a discrete-time model1, allowing to perform maintenance only within defined points in time. The model uses the following decision-making variables:

• A variable that indicates whether a maintenance task should be performed or not.
• A variable that assigns a maintenance task to the period in which it will be carried out.

There is a limit on the total time allocated to each of production and maintenance tasks.

The model's objective function is a combination of maximising the number of maintenance tasks that can be performed within the time horizon and the earliness of these tasks. A user defines parameters to tune the importance of each of these aspects.

The trade-off between production and maintenance remains unanswered

The two-step approach showcased in the paper leaves a core question unanswered; the trade off between production and maintenance. Typically, production scheduling aims to optimise a particular KPI, such as cycle time or throughput. By treating the production schedule as fixed and optimising the number and earliness of maintenance operations around it, we are ignoring the trade-off to the KPIs that matter and there may well be foregone synergy opportunities.

The paper rightly highlights the need to consider maintenance using a formal mathematical framework. Nevertheless, there are some assumptions that limit the applicability and benefit of the proposed approach.

• One limitation is the discrete representation of time; continuous-time modelling* would instead allow for a more precise indication of when the various tasks should be carried out. This is especially relevant for modelling more complex tools e.g. photolithography, where maintenance lasts only a fraction of the discrete time.
• The model makes use of the concept of “tool families”, to capture the fact that many tools are identical, which allows for substantial model simplification. However, in practice, most tools will have individual characteristics (such as the set of recipes they can carry out, or the secondary resources they may be using), which renders them non-interchangeable.
• Another consequence, mentioned also by the authors, is that all maintenance is treated as optional, with no possibility to mark a particular task as “must-run” since that may require amendments to the original production schedule for feasibility.

Want to know more about the different wafer fab scheduling approaches including heuristics, mathematical and hybrid approaches? Read the following whitepaper, here we cover everything about wafer fab scheduling approaches

Business perspective

Results are presented for 12 real-world case studies, involving around 100 maintenance tasks to be scheduled for 14 tools families over a span of 60 one-hour periods. The more relaxed model (model 2) is shown to perform better, both in terms of the number of maintenance tasks planned as well as total earliness. One strength of the proposed approach is the speed of computation. As the authors state, the proposed model can be the basis for iterative discussions between production and maintenance planners.

Flexciton’s view on solving wafer fab challenges

Ideally, production and maintenance scheduling should be tackled in a single model, where the objective function according to cycle time and throughput applies. This can be achieved by treating maintenance as tasks that need to be scheduled within a specific window. Thereby fab managers can explicitly consider the impact that maintenance tasks have on the schedule such that impact on the production KPIs is minimised. Of course, such integrated approaches result in a substantial increase of problem size and complexity, necessitating the development of solution strategies capable of handling the ensuing complexity. Especially in cases of a large number of maintenance tasks or lengthy maintenance, such constraints can quickly render a problem intractable.

Flexciton offer wafer fabs smart scheduling solution

At Flexciton, we have developed smart scheduling solution that involves decomposition techniques to manage the added complexity introduced by maintenance constraints. The users can describe their maintenance tasks as “optional” or “must-run” as well as having a fixed start time or a flexible time window within which they can be carried out.

The Flexciton engine proceeds to optimise the target production KPIs while respecting maintenance constraints. The resulting production schedule prescribes the best time to carry out maintenance while capturing all individual tools characteristics and respecting all operational constraints so as to achieve the best use of available assets. Learn more about the technology behind Flexciton's smart scheduling solution

* Discrete time and continuous time are two alternative frameworks within which to model variables that evolve over time. Discrete time views values of variables as occurring at distinct, separate "points in time". In contrast, continuous time views variables as having a particular value for potentially only an infinitesimally short amount of time. Between any two points in time there are an infinite number of other points in time

‍

‍

Insightful experiments expose the weakness of limiting the number of recipes enabled on a tool. The key findings are that this limitation can lead to an increase in fab cycle times by more than 40 percent.

It’s not easy managing a fab. While the goals are simple – maximising the yield and throughput – execution is really hard. That’s partly because fabs churn out a range of products, produced within many tools and processes; and partly because the goalposts constantly shift, due to the dynamic nature of the environment.

Sometimes the need for change reflects success in the business. After winning a new order, those running a fab may need to develop and run new recipes before they can manufacture these latest products. Unfortunately, this requires manpower to implement, it could dictate the need for more regular maintenance of a tool, and evaluating KPIs could prove tricky.

At other times those that run a fab will have no warning of the need for change, and will be forced to make tough decisions at breakneck speed. If a tool suddenly fails to process material within spec it has to be taken off-line and assessed. Meanwhile compromised wafers are etched back and processed via a different route through the fab, potentially involving alternative tools running new recipes.

To simplify operations within a fab, many managers restrict some tools to processing only particular products. This is accomplished by limiting the number of recipes on selected tools. It’s a tempting option that might reduce tool maintenance, but it is not without risk. While the hope is that this course of action has negligible impact on the throughput of the fab, there is a danger that it could make a massive dent in the bottom line.

Up until now, fab managers have taken an educated guess at what the implications might be. But they would clearly prefer a more rigorous approach - and thankfully that is now within their grasp, due to the recent launch of our Optimization Scheduler.

To illustrate some of the powerful insights that can be garnered with our software, we have considered the consequences of restricting the use of tools within a hypothetical fab. Our key findings are that this can lead to a massive hike in waiting times at particular tools, and ultimately increase fab cycle times by more than 40 percent.

Complementary case studies

We reached these conclusions after performing a pair of complementary case studies. The first, considering a single randomly generated dataset, allowed us to take a deep dive into the frequency of use of particular tools and their corresponding wait times. The second, involving twenty randomly-generated datasets, allowed us to evaluate the impact of restricting the use of tools on the cycle times for the fab.

In the fab that we modelled there were six toolsets, each with four tools. All ran until the fab had carried out what we describe as 1000 work units – that is, the fab operated until it clocked up 1,000 steps across all lots through the modelled tools (one lot has one work unit for every step completed along its route).

For our first case study, which considered a single randomly-generated dataset, we distributed the work units between the toolsets in the following manner:

The objective of this case study was to examine how, if we were to vary the number of recipes enabled on tools, the corresponding wait times at tools would change. We simulated 4 different scenarios. In the most restrictive scenario, a work unit only had the option to be assigned to 1 tool (because only that tool had the recipe enabled). In the opposite, most generous scenario, a work unit could be assigned to any of the 4 tools within each toolset (because every tool has the corresponding recipe enabled).

Clearly, restricting the number of tools has unwanted consequences on waiting times. For all six types of tool that we considered, a decline in recipe availability increased the total wait time. This is a non-linear relationship, with by far the greatest difference in wait time found when availability shifted from two tools to just one. The impact of having just one tool available is also tool dependent. For the six types of tool considered, the increase in wait time over baseline varied from less than 250 percent to just over 700 percent.

Of course, wait time is only a part of overall cycle time. To investigate the impact of tool availability on total fab cycle times, we considered twenty datasets, each with a slightly different distribution of work units allocated to toolsets. For this investigation we maintained our requirement for 1000 work units, undertaken by six toolsets, each with four tools.  

For this simulation, we flexed the recipe availability and calculated the change in cycle time. We found that when the tools were available and capable of running all recipes, flexibility was high as it could be, allowing the fab to run at its full throughput. Reducing tool availability by limiting recipes led to a significant increase in cycle time.

Plotted in the graph below are increases in cycle time resulting from a reduction in tool availability. These values are relative to the theoretical minimum cycle time, realised when all four tools are available and flexibility maximised. While there is a variation in impact across the 20 datasets, the trend is clear: when tool availability reduces, cycle time takes a significant hit. Averaging the results across all datasets (depicted by the bold line) shows that when the proportion of tools available falls to 25%, cycle time lengthens by more than 40%.

Conclusion

This pair of case studies, carried out with our smart scheduling software, has uncovered valuable lessons. While simplifying production by limiting the recipes run on toolsets may be tempting, that can cost a fab significant increase in cycle times. By utilising the "what-if" capabilities of our Optimization Scheduler, fab managers can run different scenarios for data-driven decision making and ultimately become more informed about the impact of their choices on the shop floor.

‍

Building and maintaining any form of scheduling solution to be flexible yet robust is not an easy undertaking. Commonly, fab managers have resorted to rule-based dispatch systems or other discrete-event simulation software to estimate how their fab will play out in the near future. Often this requires deciding a specific KPI that is important to the fab up-front; do I care more about getting wafers out the door, or reducing the cycle time of those wafers?

Competing objectives challenge

As a fab manager, there are a number of competing objectives to balance on the shop floor that all impact the profitability of the fab. Whether that be reliably delivering to customers their contractual quantities on time, or ensuring that fab research and development iteration time is kept low, fabs need a flexible, configurable scheduling solution that can produce a variety of schedules which account for these tradeoffs. At Flexciton, we call this “multi-objective” scheduling; optimizing the factory plan whilst considering several independent KPIs that, in this case, are fundamentally at odds with one another. This article explores Flexciton’s approach to multi-objective scheduling and how we expose simple configurations to the fab manager, whilst allowing our scheduling engine to ultimately decide on how that configuration plays out in the fab.

If there is no automated real-time dispatch system in the fab, determining the "best" schedule is a very complex procedure that cannot even be accomplished with advanced spreadsheet models. Assuming that the fab is advanced enough such that a dispatch system is in place, it will likely only consider "local" decisions pertaining to the lots that are immediately available to the dispatch system at the time the decision is made.

Dispatch systems typically do not have the configurability to adjust the user's incremental utility with respect to throughput and cycle time; they typically adhere to a series or hierarchy of rules that are tuned to consider exactly one KPI. Therefore to change the objective of the dispatch system would require rewriting these rules; an often time-consuming exercise that requires advanced technical knowledge of the dispatch system. This makes it almost impossible or otherwise very time consuming to trial various configurations of the fab manager’s preferences.

Balancing various objectives for best results

The Flexciton optimization engine is a multi-objective solution that can linearly balance various KPIs according to user-chosen weights. As these weights are exposed to the end-user, this renders the possibility of running many different scenarios with varying preferences trivial. Fab managers can have access to the specific weight values themselves or work with our expert optimization engineers to select from a handful of high-level configurations and the solution will select appropriate weights itself.

To properly understand the flexibility of the engine, we will now step through four case studies. The goal is to compare how, given the same dataset, slightly different objective configurations impact the solution that is returned by accounting for the change in preferences.

We present a schedule of nine tools from across five toolsets with seventy lots of a mix of 65% Priority1 lots. Each lot can go to a random subset of tools within a single toolset.

The schedule will then be tested against four runs:

1. Produced by a dispatch system with heuristic rules
2. Optimized for cycle time
3. Optimized for the on-time delivery of wafers
4. Balanced optimization considering both cycle time and OTD
   ‍

For each of these scenarios, we will present two gantt charts; one labelled with the “Queueing Time” of each lot (aka “rack time”) and another labelled with the “Late Time” of each lot. Late time refers to the duration by which the lot completed processing after its due date. If it was not late, the label reads “0s” since we do not consider being more early as being more favourable. Lots that are considered high priority (Priority 1 to 3) are given a circle badge indicating such. Low priority lots are Priority 4 through 10. Each lot is coloured according to this priority class.

Case study #1: base case - greedy dispatch

To begin, we’ll present how a schedule could look when produced by a dispatch heuristic that does not consider the future arrivals of wafers, but simply what is currently available in front of a tool. The greedy rule here is to just dispatch the highest priority wafer on the rack at the point the tool is idle.

In the above example, the high-priority wafers have to wait due to the system only considering what’s on the rack and therefore dispatching the low-priority wafers that are ready to go.

It should be noted that such a strategy is great for improving overall throughput and cycle time since the machine idle time is reduced by constantly dispatching wafers. This has the side effect of delivering all-bar-one of the wafers on time. In reality though, not all lots are equal and fab managers care a great deal more about certain high-priority lots thus making the scheduling problem quite a bit trickier.

Unfortunately, in order to reconfigure the system to place greater importance upon the high-priority wafers and dispatch them first would require complex rewriting of the dispatch  rules to “look ahead” at the wafers that are not yet on the rack, and are arriving shortly. The dispatcher would then elect to keep the machine idle in order to reduce the high-priority wafer cycle time.

Case study #2: Optimize for high-priority-lot cycle time

Instead of modifying the RTD rules, we can emulate what that would look like by running our optimization engine whilst optimizing for the cycle time of high-priority lots:

The low priority lots at the front of the schedule are replaced with high-priority lots so that they can be dispatched as soon as they arrive. These low priority lots have been pushed to the back of the schedule with non-zero rack time (since the cycle time of high priority lots matters so much more). Naturally this is at the cost of overall average cycle time which has suffered by 23% in order to improve Priority1 cycle time by 11%. Also note that on tool “SBXF/115”, our scheduling solution has pushed the Priority2 (orange) and the Priority10 (green) lots later so that the Priority1 (red) lots are rushed through with zero rack time.

Case study #3: Optimize for on-time delivery

With optimisation, there are no additional changes required to increase the flexibility of the system. We simply describe what a good schedule looks like using the multi-objective function and the optimizer does the rest. Subtle tweaks to this function will inevitably produce very different schedules. Now let’s take a look at how the schedule alters when we want to maximise solely on-time delivery.

As expected, cycle time is quite a bit worse than previously however now there are no lots delivered late. This is very similar to the original schedule produced by simple dispatch rules. The low-priority lots have been brought forward so that they are delivered on time and the cycle time of the high-priority lots suffer as a result.

Case study #4: Optimize for both

Finally, the main purpose of this article is to illustrate the ease of considering both KPIs with some relative weight simultaneously.

Note that the KPIs of cycle time and throughput are slightly worse than when that was the sole KPI being optimised. The key is that both are better than when the other KPI was being optimized. This balance is entirely in the hands of the fab manager. We maintain roughly the same cycle time of high-priority lots as when optimising for cycle time and fewer lots are late than when optimizing only cycle time.

Summary and Conclusions

This article has provided a number of ways that illustrate how optimization can be considered both more flexible and robust than heuristics that cannot effectively search the global solution space.

The engine is simple to tune due to the exposed weights and/or configurations presented to the fab manager which allow a high degree of customisation both with respect to the objective function and wafer priorities. This flexibility allows us to easily consider complex hierarchical objectives found in semiconductor manufacturing such as “optimise high-priority cycle time as long as no P1-8 lots are late” or “optimise batching efficiency (perhaps due to operator constraints) and then high-priority cycle time”. Ultimately, our solution is a market-leading scheduler that will realise true KPI improvements on your live wafer fabrication data.

Flexciton is currently offering the Fab Scheduling Audit free of charge. To enquire, please click here.

‍

Batch tools are purposefully built to process two or more lots in parallel. However, due to the complexity and volatility of the wafer fabrication environment, each day, wafer fabs are challenged to make complicated batching decisions. How to determine when to batch lots together and when better not? This is what we shall call the ‘batch or not to batch dilemma’.   

Why batch lots together in a wafer fab?

One of the most precious commodities is time, and batch tools are designed to make the best use out of time. Often, these tools have a long processing time. An example would be a diffusion furnace. Instead of waiting for 6 or 8 hours for a single lot to go through the whole process before loading the next one, batching allows the processing of multiple lots simultaneously. This sounds like an easy way to solve a complex problem. However, with batch tools, on top of a whole host of constraints to respect, there are exponentially more scheduling options available compared to a non-batch tool.

In addition, the processing times of batch tools introduce a very complex dynamic. Batch tool processing times consist of fixed time (regardless of the number of lots in a batch) and a variable time (increases with each additional lot). Because of the fixed time component (+ tool setup time) in creating a batch, there is the perception that larger batch sizes are more efficient.

A typical approach to batching decisions

At a batch tool, there are a number of decisions to be made, such as whether you process lots that are already in front of the tool or wait for more to arrive? Additionally, if the number of lots waiting exceeds the tool capacity, which lots should you batch together and process first?  

Typically, each fab decides on their batch-size policy, which will guide the batching decisions. One of the commonly used policies would be a Minimum Batch Size (MBS) policy, setting a minimum number of lots required to start processing. This could be determined by running a large number of different simulations. These simulations will determine the batch size that provides the best performance for that specific use case. Thus, an MBS heuristic rule would be created, setting a certain direction to follow for all future lots.

On the other hand, a ‘near-full’ policy would require you to wait until the batch size is as close as possible to the maximum capacity of a tool. In this case, it is possible to achieve high throughput, but it can also cause the tool to stay idle for a long time when waiting for more lots to arrive in order to satisfy the policy, which negatively impacts on overall cycle time.

Processing a batch whenever the tool is available and ready to process, is another approach (this is called the ‘greedy policy’). This may reduce cycle time during times of low WIP, but will likely cause an increase to cycle time and lower throughput during times of high WIP.

How to determine the right batch size

The fact of the matter is that there isn’t one perfect batch size. In reality, it depends on the context of the whole system at that specific point in time. The batching decision relies on a number of dynamic factors:

• Max batch size constraints
• What will arrive and when
• What is currently in the queue
• Possible recipe combinations that can be batched together
• What priorities the wafers have
• What objective you are currently optimizing for
  
  ‍

In many fabs, daily batching decisions are guided by a dispatch system which uses rules-based heuristic algorithms. This approach can work very well in some cases but can bring very poor results in others.

Let’s take a look at an example, where we use a simplistic approach to illustrate the problem.

Suppose we have a batch tool with a maximum batch size of 4 lots. In order to get better efficiency, a typical dispatch rule is to set a minimum batch size e.g. minimum batch size of 2 lots. However, in a situation where one lot is already present, waiting too long, will be inefficient. Therefore, a maximum wait time - let’s say 30 min - would apply to the rule of minimum batch size. If we have waited 30 min, and another lot did not arrive, the dispatch system would send one lot for processing.

Sometimes this works well - here another lot arrived within 5 minutes, allowing  2 lots to be processed simultaneously.

In this example, the rule was effective and achieved a lower average cycle time than if we hadn’t waited for the second lot to arrive. However, sometimes a rules-based dispatch system can make poor decisions. Typically the dispatch system will only make local decisions, and won’t look ahead to anticipate which lots and when are coming, as it’s illustrated in the second scenario below. The second lot arrives at the tool in 60 minutes, but due to the 30 min waiting rule, the first lot has already been dispatched.

In order to make more optimal decisions around batch sizes, we need to be able to anticipate what WIP will arrive, when it will arrive, and where this WIP goes next after the current batching step.  This can be achieved by applying a smart scheduling approach which understands the broader context required to make optimized batching decisions. The below example illustrates a rules-based decision vs an optimization approach.

This is one very simple example which shows some of the trade-offs that must be considered when making batch decisions. In the above case, it would be possible to write an extension to the dispatch rules to account for the scenario presented. However, in reality, there are several other factors which bring additional complications. For example, often, Lots will have different priorities. When a high priority lot is batched together with a lower priority lot, the average cycle time for both lots may be reduced compared to running the Lots in two sequential batches. That said, the cycle time of the high priority lot is likely to be increased - this may be undesirable.

Given how dynamic a fab is, writing dispatch rules to efficiently deal with the full range of scenarios is possible, but would be extremely time-consuming to maintain and expensive to build.

Conclusion

Batch tools are extremely complex machines to schedule, as there is a huge number of scheduling options, and each option has a different efficiency. Commonly used dispatch rules can cause poor performance in a dynamic fab environment. Often, the batching methodology follows a fixed rule, such as maximum batch size. These fixed rules can provide occasional good outcomes, but they are unable to consistently provide good solutions. As a result, the KPIs across the batch toolset might show undesirable increases in cycle time, or reductions in throughput if the WIP mix or objectives change. Additionally, creating very efficient rules would require a lot of time and extensive maintenance.

Smart scheduling, on the other hand, introduces the ability to make optimized batching decisions in any situation to achieve the objective of increased throughout or lower cycle times. By applying mathematical hybrid-optimization techniques, we are able to find a solution which is near-global-optimal, delivering a consistently high-quality outcome.

Get your Wafer Fab Scheduling Performance Analysis fully remotely and free of charge. Click here to get in touch.

‍

In our previous blog, we talked about The Theory of Constraints. One of the principles being, the system is only as good as its weakest module - i.e. a bottlenecked tool or toolset.

Because of bottlenecked toolsets, wafers will spend a great proportion of their cycle time queuing (non value-adding) rather than processing. The longer or more uncertain the wait time, the higher risk of variability in the cycle time. This ultimately impacts the overall productivity of a fab. ‍

The wait time challenge

The Pareto Principle (aka the 80/20 rule) implies that most things in life are not distributed evenly. The very same rule can apply to wait time vs processing time in fabs. We have carried out performance analyses of numerous fabs recently, and we have witnessed the Pareto Principle repeatedly across various wafer fabs.

It is obvious that in order to maintain high performance in a fab, wafers are required to be processed within a predictable time frame and in doing so, spend as little time as possible queuing. However, it is very common to observe a significant disproportion within cycle time, where the wait time is significantly longer to processing time. In some cases, we have seen wait/process time ratios of 70% to 30%, respectively.

If we can reduce the queue time, we can reduce cycle time. By drilling down the wait time, we repeatedly observed the Pareto Principle highlighting that ~20% of toolsets contribute to ~80% of all wait time.

The bottleneck tool is not always the root cause

In order to solve the wait time problem, we first need to identify where in the fab the wafers spend the most time waiting to be processed. Typically, the culprit toolsets are those which deal with wafer re-entry and possess challenging constraints that heighten the scheduling difficulty, such as photo, EPI and Implant.

Improving the sequencing or mix of WIP directly at the bottlenecked toolsets seems to be the obvious starting point. However, whilst the bottleneck toolsets are incurring the most amount of wait time, they are not always the root cause of the problem. This may occur elsewhere.

For example, if the toolset is being fed WIP in which the associated recipe is unavailable for the next few hours, then this WIP has no option other than to wait. The toolset becomes bottlenecked, however, the root cause was ineffective WIP management of toolsets upstream.

Optimizing WIP to reduce wait times

In order to reduce wait times at the bottleneck areas, we must control and optimize the WIP flow. This can be achieved by applying smart scheduling, which advises dispatch on the best real time decisions to take.

Smart scheduling can first be applied to the bottleneck toolset in question to drive local efficiencies that relate to the specific constraints and cost functions of that toolset. For example, you can use optimization to:

• Optimize batching decisions (the wait or not to wait dilemma)
• Optimize changeover sequencing
• Balance the toolset loading
  
  ‍
  

By optimizing at a local level, you can achieve increases in efficiency, which leads to cycle time reduction. However, the scheduling is still at the mercy of which WIP is sent from upstream toolsets. The next step is to schedule toolsets globally (scheduling multiple toolsets together) so that the WIP mix arriving from upstream toolsets is optimally sequenced to facilitate the heavily loaded bottleneck.

Referring to the example in the previous chapter, by applying advanced global scheduling, the upstream toolset can be scheduled to prioritize WIP such that it matches the available recipe of the downstream toolset. As a result, the wait time at the bottleneck toolset can be reduced even further than if you had just optimized the toolset locally. Additionally, by having a better future forecast on where WIP will be, you can better optimize load port conflicts (IPP conflicts) and auxiliary resources (reticles at photo tools).

At Flexciton, our optimization strategy is to:

• Optimize local toolsets which incur most cycle time
• Expand to global scheduling to balance the entire system
  ‍

Identifying the problem

By taking a global view of a fab, we can analyse the cycle time across the whole fab and begin to determine where to focus optimization efforts in order to improve cycle time KPIs.

At Flexciton, we have developed an analytical tool (Wafer Fab Performance Analysis), which uses historical MES data to provide a comprehensive view of fab performance. Using the tool, we are able to locate bottleneck toolsets across the fab and quantify the impact each toolset has on cycle time. Through this global analysis, we are able to recommend an optimization strategy outlining where cycle time improvements can be made.

Although fab managers are acutely aware of the overall efficiency of cycle time at their fab, employing the Wafer Fab Performance Analysis allowed us to accurately quantify the problem and pinpoint where efficiencies are lost.

Click here to get in touch and learn more about complimentary Wafer Fab Performance Analysis for your fab.

‍

Any manageable system is limited by at least one constraint. What if this system is the most complex manufacturing process in existence? Providing optimized production scheduling for wafer fabs, we deal with a great number of constraints.

In the early 80s, Eli Goldratt, a physicist turned management guru published his book "The Goal". Why is it important? In the book, he underlined the foundations for what was described later as the "Theory of Constraints".  

Although it was written nearly 40 years ago, and manufacturing industries have since evolved, many of the principles stated in the book remain relevant today. They also mirror what we do at Flexciton, helping wafer fabs to improve performance by optimizing the way the fabrication process is scheduled.

Let us give you a snippet of what is it about by sharing the three key messages of the Theory of Constraints:

• "Throughput is the money coming in." Balancing the flow of product through the plant with demand from the market is critical for the business performance. Throughput is usually a priority KPI we follow to calibrate our scheduling solution for a given fab.  
• "Every hour lost at a bottleneck is an hour lost in the entire system." Based on the Wafer Fab Performance Analysis which we have carried out for various fabs, we see that a small number of tools are bottlenecks, yet this seemingly little number are responsible for substantial wait time in the fab.
• "We shouldn't be looking at each local area and trying to trim it. We should be trying to optimize the whole system. A system of local optimums is not an optimum system at all." At Flexciton, we first take a global look across the whole fab, then we identify areas where optimization would bring the most significant improvement for the entire facility.
  
  ‍

The very first step for improving the fab productivity and KPIs such as throughput or cycle time is to identify the system's constraints and where the efficiency is lost. Flexciton's Wafer Fab Performance Analysis takes a global view of a fab and pinpoints the areas or specific toolsets responsible for lower productivity. This analysis often brings unexpected and eye-opening results.

We are currently running the Wafer Fab Performance Analysis fully remotely and free of charge. If you wish to discover more, please click here to get in touch with one of our consultants.

‍

A consistent metric that wafer fab and foundry managers seek is high throughput. Being able to control and maximise throughput is critically important to the health and profitability of a semiconductor business. If the factory in question is capacity constrained, then any percentage increase to total fab throughput can be converted into further revenue for the business. Higher throughput can also prevent the need to invest in additional expensive CAPEX.

Achieving operational excellence in semiconductor manufacturing is a very difficult task and requires sophisticated industrial engineering for the best results. Fabs are highly dynamic and can be unpredictable at the best of times (tool downtimes and changing priorities), thus sustaining consistently high throughput is a major challenge for any fab manager or industrial engineer.

In this article, we are going to outline two factors that are key culprits in limiting semiconductor fab throughput, and how to control them by applying smart scheduling strategy.

Factor #1 - WIP mix at bottleneck toolsets

Wafer fabs have to deal with managing bottleneck toolsets since tools can be expensive and some deal with reentrant flow. Bottlenecks have a negative effect on cycle time and overall throughput, however, it is how we manage the sequencing and flow of WIP through these bottlenecks that leads to a productive and well-balanced factory.

At Flexciton, we repeatedly see Pareto's principle reveal that 20% of toolsets are responsible for 80% of wait time in the fab. Tackling the queue time first behind these key toolsets can significantly help with maximising global throughput across the fab, and should be on the priority list for industrial engineers.

Production progresses as quickly as the worst bottleneck allows. An hour lost at a bottleneck is an hour lost for the entire system (following the theory of constraints).

When a toolset is a fixed bottleneck (always in high demand), then it should be a priority to increase throughput at this step as the first point of call. Thus, the rest of the fab benefits from a higher overall throughput as a result. This can be achieved by improving load balancing across the toolset and better scheduling of upstream WIP to ensure optimum sequencing. For more complicated toolsets, you may need to consider optimizing changeovers or consider the scheduling of auxiliary resources (reticles for photo) in order to provide relief to the bottleneck.

It is also common to witness dynamic bottlenecking, whereby the bottleneck toolset may change over time as a result of the changing factory dynamics. For example, WIP might suddenly inflate at implantation toolsets from time to time, but will not always have a consistently high level of WIP behind it. Here, managing the global flow of WIP between toolsets (upstream and downstream) helps to alleviate these dynamic cases of WIP buildup, leading us on to the second discussion point below.

Factor #2 - Local optimization of toolsets

It is common to measure fab performance using local KPIs and targets, broken down by area or toolset for the operators. Thus, there is pressure to maintain constant high throughput across all toolsets (with little or no consideration of the impact this has on upstream or downstream tools).

Unfortunately, when local optimization occurs within toolsets, it can result in a significant imbalance to the overall fab system - leading to poor throughput. Put simply, if we process a load of WIP through 'toolset A' really quickly, then this may end up sitting at 'toolset B' for hours because the recipes are not available! Going back to point 1, this is highly inefficient and is likely to inflate bottlenecks and grind away at your throughput potential!

A wafer fab is a complex, intricate system, and we must optimize the flow across all toolsets to ensure high overall throughput. A fab in which all tools are working at max capacity at all times is actually very inefficient, and a system of local optimums is not an optimum system at all.

We want to ensure that what is being produced upstream of bottleneck toolsets are fed with optimum WIP mixture. Doing this can provide exponential increases in a fab overall system efficiency and ultimate throughput!

Smart scheduling, and its impact on throughput

The key to managing WIP flow across the entire fab system is to introduce high-quality scheduling into the process. By being able to comprehend upstream and downstream WIP flow, we begin to schedule globally across toolsets (rather than locally), and this is the key to minimising the severity of dynamic bottlenecks.

Smart scheduling is critical, as it allows the fab manager to balance priorities and make smart decisions in advance that aid greater throughput. By feeding directly to the dispatch system or providing operators with clear, specific direction on the next actions, it improves predictability and performance throughout the fab.

‍

Scheduling is difficult. Scheduling your workday ahead in the morning is difficult. However, this blog is about wafer fabs, and let's start by saying that scheduling a wafer fab to run optimally is one of the most challenging mathematical problems that exists in modern day manufacturing. Why?

Wafer fabrication is extremely complex. Lots of wafers may re-enter tools numerous times, and compete for capacity against hundreds of other lots. Not to mention a whole host of unique scheduling constraints as well as unexpected events, such equipment suddenly going offline.

Scheduling problems are widely researched, and there are various methods used to solve them. However, in the semiconductor industry, there are two commonly debated techniques: heuristics (rules-based) and mathematical optimization.  

Both methods provide effective scheduling solutions, however, each has limitations, which is the topic of the discussion below.

Heuristics

A heuristic algorithm is essentially a ‘best-guess’ approach to a decision problem.  

By taking approximations and algorithmic shortcuts, heuristics can arrive at a final decision very quickly. Think of a heuristic as a long decision-tree of if-then logic. Heuristics are fantastic at finding realistic solutions extremely quickly. Since wafer fabs are extremely dynamic, this method is widely adopted for this reason.

However, the benefit that heuristics have on speed, gives way to its biggest flaw - solution quality. Being rules-based, it can only search through a restricted number of scenarios and follow a 'familiar' path. Hence, you never really know how good the final solution is in terms of quality (there could be a much more optimal solution available).

A heuristic simply cannot optimize a decision problem - leaving wasted productivity on the table.

Mathematical optimization

Mathematical Optimization is used when you want to find the optimal (best possible) solution.

A classic optimization problem is the ‘travelling salesman’ problem, where the objective is to seek the shortest path for a salesman to travel, given a number of nodes that must be visited. Here there is only one optimal solution, and mathematical optimization is a method of finding it.  

In wafer fab scheduling, optimization can be used to similar effect: where, as an example, the objective may be to find a schedule that minimises cycle time for both production and R&D wafers. Here, all possible scheduling scenarios can be evaluated and the best one chosen (a major benefit over heuristics, where only one schedule is evaluated and chosen).

In a fab, factory states change frequently making it an extremely dynamic environment to schedule. As a result, a scheduler must be fast enough to re-run schedules in order to cope with the factory dynamics. However, optimization algorithms need time to find the best possible solution. In practice – it might be even hours, which makes pure optimization scheduling impractical and unrealistic for wafer fabs.

Which method for high-quality wafer fab scheduling?

This is the dilemma. Heuristics are widely adopted because they are fast and realistic. But the quality of solutions is inferior compared to the potential of using an optimization technique. Pure optimization, on the other hand, is not feasible in a fast-paced live fab environment.

What is needed is a ‘best of both worlds’ approach. A scheduler that is quick, reliable and optimal. This is exactly what a hybrid-approach seeks to achieve.

In our recently published white paper, we examine a new hybrid optimization approach that enables fast and near-optimal scheduling. In this, we compare the hybrid approach with the two widely used methods described above. The white paper can be viewed and downloaded by clicking here.