Decision-Making Techniques for Smart Semiconductor Manufacturing

In the opening session, the organizers welcomed the participants. Next, the participants each introduced themselves. An overview of the goals and objectives of the seminar and a detailed review of the seminar program including the ground rules for interactions followed after the introduction. The remainder of the day on Monday consisted of an introduction into smart manufacturing (by John Fowler and Lars Mönch) and six industry overview talks (by Adar Kalir, Marcel Stehli, Alexandru Prisacaru, Thomas Ponsignon, Hans Ehm, and Peter Lendermann). Tuesday and half a day on Wednesday were devoted to presentations and discussions about the various elements of the semiconductor supply chain planning and control systems shown in Figure 1 above and their relations to smart semiconductor manufacturing. See Table 1 below for a list of topics and presenters and Appendix B for abstracts of the presentations.

Wednesday afternoon was the excursion to Trier that was enjoyed by the participants.

Thursday was devoted to a set of three breakout sessions with report outs on the topics in Table 2. Appendix C has the breakout report outs.

The first set of breakout sessions had three groups focus on machine learning (ML) since ML techniques are an important element of smart manufacturing. The second set of breakouts had again three groups consider simulation-based decision support since such techniques are an important element of smart manufacturing. The third breakout session had again three groups, one dealing with sustainability issues as a core element of smart manufacturing, one with information systems including ontologies, and a last one that prepared a joint session with the participants of the Dagstuhl Seminar running in parallel with the present seminar. During the joint session with the other seminar on Thursday evening we learned from some basic facts about multi-criteria optimization (well-known for many participants of our seminar) and introduced various multi-criteria optimization problems found in semiconductor manufacturing to the participants of the other seminar.

Friday consisted of a panel discussion (panelist Heavey, Lendermann, Matta, Pedrielli) moderated by John Fowler on the required core elements of a simulation testbed which allows for assessing smart planning and control decisions in the semiconductor industry and a wrap-up session.

There were a number of key findings and areas for future research that were identified in the seminar. We will first summarize some of the key findings and will follow this with some areas for future research.

One of the first findings was that the participants generally agreed that some of the major elements of smart manufacturing are already implemented in semiconductor manufacturing, but there are also elements that are less well understood and consequently implemented. Having said this, ML approaches are considered as promising for semiconductor manufacturing, but their potential is still not fully understood and explored. This is especially true for the role of reinforcement learning which is recently often applied to semiconductor scheduling.

Second, it appears that there are still limitations in applying different simulation paradigms in practice such as ML approaches are often not integrated into simulation models and existing reference simulation models are too difficult to apply for benchmarking purposes. Digital twins are considered as another promising direction for semiconductor manufacturing, however, they are not fully implemented so far.

Third, both the industrial and academic participants generally agree that the integration of sustainability efforts into decisions on the wafer fab and the supply chain level is often fairly ad hoc and could/should be improved in the future.

Finally, the participants generally agreed that there does not currently exist an adequate reference (simulation) model for smart manufacturing. Such a simulation model should allow for making sustainable decisions and for supporting the application of ML techniques when decisions are made.

In addition to the findings mentioned above, several areas for future research were identified.

An overarching idea was that the future research should focus more on formulation of appropriate models for smart manufacturing because this is fundamentally more important than the actual solution techniques chosen. Some of the future research areas are included below:

• $\mathrm{■}$

  More applications of CPSs, for instance for AMHS operations or for lot processing are desirable in semiconductor manufacturing.

• $\mathrm{■}$

  Multi-agent systems (MAS) are a desirable software paradigm for smart semiconductor manufacturing. However, more real-world applications are required.

• $\mathrm{■}$

  Developing better integration of various decisions made in the elements of Figure 1.

• $\mathrm{■}$

  Incorporating sustainability aspects into strategic and tactical supply chain planning models.

• $\mathrm{■}$

  Exploring the use of different simulation paradigms (systems dynamics, agent-based, hybrid models, reduced simulation models) to model and analyze semiconductor supply chains.

• $\mathrm{■}$

  Exploring and applying the possibilities of the semantic web to facilitate a meaningful data exchange between different planning and control applications.

As a way to further the discussion of and collaboration on the topics of the seminar, Prof. Lars Mönch, Hans Ehm, and Prof. John Fowler are guest editing a special issue of the Flexible Services and Manufacturing Journal entitled Decision-Making Techniques for Smart Semiconductor Manufacturing. The deadline for submission is May 1, 2024. This date was selected to allow time for ideas created by the participants of the seminar to be incorporated into papers https://www.springer.com/journal/10696/updates/26269410.

First, we want to thank the Dagstuhl staff for their great support of this seminar. The seminar also would not have been nearly as productive without the active contribution of every attendee, and for that the organizers are extremely grateful.

The presentation focused on the flexible job-shop scheduling problem and some of its extensions, on neighborhood-based metaheuristics and their application to semiconductor manufacturing. The extensions include additional constraints, such as sequence-dependent setup times and batching, and new criteria. The disjunctive graph modeling was presented for various complex flexible job-shop scheduling problems, together with various properties that allow to speed up the search in neighborhhod-based metaheuristics. Scheduling in semiconductor manufacturing was then discussed to emphasize the complexity and the size of the real-life problems that must be solved. A batch-oblivous approach was then presented, which has been implemented and is being used in a real factory to solve problems with more than 2,500 operations and 200 machines. The presentation ended with some general conclusions, in particular on various industrial constraints not discussed but already considered, and perspectives on future relevant academic and industrial research, in particular robust scheduling.

Corona triggered a bullwhip amplified demand reduction for semiconductors – especially from the automotive industry. In conjunction with rising demand in communication industry (more home offices, more cameras, more audio tools, …) global capacities for the automotive industry was lost and this caused global chip shortages when the demand came back in the automotive industry with the consequence of shutting off of car manufacturing factories.

With analytics and simulation the root causes of the root causes could be identified, which is beyond others a Kanban driven replenishment which acts a bullwhip accelerator in disruptive times like during COVID and the human behavior based on prospect theory.

Due to the magnitude and possible further impacts of the problem not only the usual business stakeholders searched for long term solution also governments got involved and triggered and supported decision techniques to learn from the problem and mitigate it for the future.

The 4-step plan emerged: 1) higher inventory, 2) anonymous survey and 3) breakdown this coarse survey results with 3) semantic web based AI techniques to enable a 4) a leadtime based pricing. This plan is now in implementation in EU funded projects and supported by semi IAC

Furthermore general current problems and opportunities in the domain semiconductor and supply chains containing semiconductors have been shared from an industry point of view.

In the URL noted above full text of around 100 papers relevant for the domain (not limited to the presentation title), are provided.

Power semiconductor components have to fulfill high-quality standards. To meet these requirements, so-called Reliability Product Testing (RPT) laboratories perform product qualification tests, process monitoring tests, and tests accompanying the technology development. Reliability testing is resource-intensive, requiring trained engineers and high-tech equipment. The complex allocation of tests to the respective resources to create a scheduling plan within RPT labs is a very challenging task. Currently, this plan is created by senior experts with the help of a static dispatcher that doesn’t consider resource capacities. Introducing scheduling models to the RPT labs has the potential to reduce equipment idle times, to avoid bottlenecks, to meet the costomer deadlines more confident, and to use resources more efficiently. Also, costly lab extensions and equipment purchases can be prevented. The problem we are considering belongs to the family of complex job shop scheduling problems, having the following $\alpha |\beta |\gamma$-representation:

$F{J}_m|aux,prec,reentr,w_j,r_j,d_{ji},t_{jik},p-batch,B_k,s_j,s_f,pmtn,s-batch,s_{ji\mathrm{\ell }k},incompatible|\sum w_j{T}_{ji}+\sum w_j{T}_{jik}+\alpha \cdot C_{max}.$

The development of a rolling-horizon approach to solve the problem using metaheuristics based on disjunctive graphs will be investigated.

Simulation Optimization (SO) techniques refer to a set of methods that have been applied to stochastic optimization problems, structured so that the optimizer(s) are integrated with simulation experiments. Although SO techniques provide promising solutions for large and complex stochastic problems, the simulation model execution is potentially expensive in terms of computation time. Thus, the overall purpose of this research is to advance the evolutionary SO methods literature by researching the use of metamodeling within these techniques. Accordingly, we present a new Evolutionary Learning Based Simulation Optimization (ELBSO) method embedded within Ordinal Optimization. In ELBSO a Machine Learning (ML) based simulation metamodel is created using Genetic Programming (GP) to replace simulation experiments aimed at reducing computation. ELBSO is evaluated on a Stochastic Job Shop Scheduling Problem (SJSSP), which is a well known complex production planning problem in most industries such as semiconductor manufacturing. To build the metamodel from SJSSP instances that replace simulation replications, we employ a novel training vector to train GP. This then is integrated into an evolutionary two-phased Ordinal Optimization approach to optimize an SJSSP which forms the ELBSO method. Using a variety of experimental SJSSP instances, ELBSO is compared with evolutionary optimization methods from the literature and typical dispatching rules. Our findings include the superiority of ELBSO over all other algorithms in terms of the quality of solutions and computation time. Furthermore, we present how approaches similar to the ELBSO method could be integrated with a Manufacturing Execution System (MES) in semiconductor manufacturing to allow scheduling at an operational level.

The design and the implementation of a multi-agent system (MAS) for a borderless fab scenario is presented. In a borderless fab scenario, lots are transferred from one semiconductor wafer fabrication facility (wafer fab) to another nearby wafer fab to process certain process steps of the transferred lots. Production planning is carried out individually for each of the wafer fabs. The modeling of the available and requested capacity in the production planning models of the participating wafer fabs is affected by the lot transfer. Three scenarios, namely no borderless fab (NBF), borderless fab with no production planning (BF-NPP), and borderless fab with advanced production planning (BF-APP) are discussed and the performance results are presented. The transfer of the route information from one wafer fab to another one to automatically generate the linear programming models for production planning is described. Production planning is carried out in a rolling horizon setting. We show by simulation experiments that a correct modeling of the capacity in the production planning formulations results in improved profit compared to a setting where the lot transfer is not taken into account in the planning formulations. In addition, we demonstrate that an ontology to standardize the data exchange between the wafer fabs can be beneficial in a borderless fab setting.

This presentation delves into the evolving realm of automation, emphasizing the role of robot collaboration intelligence. Leveraging advanced AI, this innovation facilitates synergistic interactions among industrial robots, streamlining their control and management. With the rise of adaptable agent-based robots, such as automated guided vehicles (AGVs), autonomous mobile robots (AMRs) and overhead hoist transporters (OHTs) in manufacturing, there’s a marked increase in operational flexibility. As these robots grow more sophisticated and their numbers expand, robot collaboration intelligence emerges as a pivotal tool, amplifying their efficacy. Through industry case studies, we will elucidate the immense potential of this nascent technology:

• $\mathrm{■}$

  Management of over 1,000 Overhead Hoist Transport (OHT) Systems in semiconductor fabrication plants

• $\mathrm{■}$

  Fleet regulation of 200 AGV/AMR in warehouse settings

Additionally, we introduce the Digital Twin concept tailored for Robot Collaboration Intelligence. The outlined Digital Twin (OMS-DT) encompasses a robot emulator, integrated hardware-software links, and a simulated environment, with real-world applications of this Digital Twin being showcased.

Semiconductor manufacturing is data-intensive. It is also very complex. Combined, these two aspects drive ML usage in this industry. Big success is already evident in Yield, Equipment Diagnosis, … and growing in productivity, capacity. “Signal-to-Noise” is still a challenge in many problems (e.g. wafer breaks; CQT’s [violations]).

We introduced the “multiple orders per job” scheduling problem (Mason et al., 2004) in a two-stage permutation flowshop with the goal of minimizing the makespan. This problem, $F2\parallel moj(.),Bk(.),prmu\parallel C_{max}$ is NP-hard. We discussed different types of methods we used to solve this problem for both small-sized and large-sized problem instances. These methods include exact methods, heuristics, and metaheuristics. Finally, we compared the performance of these methods on the different problem instances that were generated using a full factorial numerical experiment.

After initial discussion a number of questions were derived among the breakout group. These questions are given below:

What are the problems that ML can tackle in semiconductor topics:

• $\mathrm{■}$

  Root cause analysis: take measures for specific lots and see the impact on the yield of the different parts

• $\mathrm{■}$

  Many applications are now in the control part, examples are:

  • –

    Sampling of the lot: how do one choose if a lot is important to measure a lot or not

  • –

    Process control

    • *

      Virtual metrology:

      • ·

        Predict the value of the measure. Try to predict the value of the metric

      • ·

        measurements

      • ·

        SPC enhancement, with techniques that replace the traditional SPC with control charts

      • ·

        in general advanced process control (APC)

      • ·

        equipment Process Control: controlling the machining parameters

      • ·

        run-to-run: take measures and change machining parameters and ML is used to build the run-to-run model

  • –

    Production control

  Cycle time prediction is another important area: ML is paired with the MES system in this case. It could be:

  • –

    Fab level:

  • –

    Single step:

  • –

    Phases in the fab that predict where the lot is going to be in the next few days: this can be for:

    • *

      Production planning

    • *

      Scheduling

• $\mathrm{■}$

  Dispatching & Scheduling (ML is used more towards dispatching)

• $\mathrm{■}$

  Lot release/order release could also be done via ML

• $\mathrm{■}$

  ML to shorten simulation time, through the use of a metamodel

• $\mathrm{■}$

  Optimize parameters of a dispatching rule to calibrate the hyperparameters of the policy

• $\mathrm{■}$

  Predictive Maintenance:

  • –

    ML can be used to predict the failure time

  • –

    Policy improvement can be used in the context of Reinforcement Learning

• $\mathrm{■}$

  Optimizing transportation within the material handling systems:

  • –

    Evaluation of vehicle health

  • –

    Routing of vehicles in the system

Simulation-based decision support remains an integral tool for semiconductor manufacturing optimization. Using simulation-based methods can enhance operational efficiency, from individual tools to entire supply chains, fostering adaptability and strategic decision-making. Addressing current challenges in the field will help semiconductor manufacturers to improve their operations and align them with business goals, in this fast progressing industry.