This is a panel paper which discusses the use of discrete-event simulation to address problems in semiconductor manufacturing. We have gathered a group of expert semiconductor researchers and practitioners from around the world who (often successfully) applied discrete-event simulation to semiconductor-related problems in the past. The paper collects their answers to an initial set of questions. These serve to not only showcase the current state-of-the-art of discrete-event simulation (DES) in semiconductor manufacturing but also provide insights into where the field is heading and the impact it will have on our world by 2050 in the semiconductor manufacturing domain.

This paper introduces the first fully integrated algorithm for combining simulation with numerical linear algebra, as a means of computing stationary distributions for Markov chains and Markov jump processes. We use linear algebra to analyze the “center” of the state space, while simulation is used to estimate contributions to the steady-state from path excursions outside the “center”. The method yields consistent estimators for stationary expectations, and can be viewed as an application of the variance reduction technique known as conditional Monte Carlo.

For large, complex simulation models, simulation metamodeling is crucial for enabling simulation-based-optimization under uncertainty in operational settings where results are needed quickly. We enhance simulation metamodeling in two important ways. First, we use graph neural networks (GrNN) to allow the graphical structure of a simulation model to be treated as a metamodel input parameter that can be varied along with real-valued and integer-ordered inputs. Second, we combine GrNNs with generative neural networks so that a metamodel can rapidly produce not only a summary statistic like E [ Y ] , but also a sequence of i.i.d. samples of Y or even a stochastic process that mimics dynamic simulation outputs. Thus a single metamodel can be used to estimate multiple statistics for multiple performance measures. Our metamodels can potentially serve as surrogate models in digital-twin settings. Preliminary experiments indicate the promise of our approach.

Call center agent scheduling is the process of assigning agents to their respective shifts throughout a day in which information regarding the volume and arrival profile of calls is unknown. The construction of such a schedule will have a direct impact on the quality of service and the finances of this call center. Of great importance is knowing when and how to assess this agent schedule despite the uncertainties of how the day will actually unfold. The answers to these questions and their contribution to maintaining a certain level of performance are explored. Through discrete-event simulation, we were able to simulate different agent schedules of a call center for the disabled community, anonymized as the abbreviation ANGUS. Our results indicate the capability of evaluating schedules based on the simulation’s predicted outcomes. With such insight, it is indeed possible to meet the performance criteria objectives developed by ANGUS.

With the recent increased pressure on supply chains, Automated Storage and Retrieval Systems (AS/RS) play an key role in improving the efficiency of logistics processes. Data-driven simulation modelling coupled with Machine Learning (ML) algorithms might represents an effective approach to anticipate problems that may occur in the warehousing processes. Thus, this PhD project aims to develop an advanced data-driven simulation model for supporting the decision-making process related to AS/RS operations. To this end, the objectives are to evaluate the state-of-art of AS/RS simulation, develop and validate a data-driven simulation model, and integrate it with a ML algorithm to reduce the cycle time. Finally, potential application of this approach will be tested in different operational settings.

With the deployment of more software in vehicles, the need for efficient and cost-effective updates arises. One method to face this need is cooperative downloading, which enables vehicles to exchange parts of an update with each other, to collect a full set. Thereby, simulation of applicable strategies is crucial, as these can influence the efficiency of such systems, however, state of the art network simulators are not build for large scale test scenarios with long simulated time-spans. We introduce Cooperative Downloading in Python (CoDiPy), a framework for efficient analysis of cooperative downloading, while improving execution time. With CoDiPy, we have investigated relevant topics like encoding, communication strategies and cost optimization.

Memory system is critical to architecture design which can significantly impact application performance. Concurrent Average Memory Access Time (C-AMAT) is a model for analyzing and optimizing memory system performance using a recursive definition of the memory access latency along the memory hierarchy. The original C-AMAT model, however, does not provide the necessary granularity and flexibility for handling modern memory architectures with heterogeneous memory technologies and diverse system topology. We propose to augment C-AMAT to take into consideration the idiosyncrasies of individual cache/memory components as well as their topological arrangement in the memory architecture design. Through trace-based simulation, we validate the augmented model and examine the memory system performance with insight unavailable using the original C-AMAT model.

Digital Twins have recently emerged as a major new area of innovation. Digital Twins are often found at the core of “smart” solutions that have also emerged as major areas of innovation. Modeling and Simulation (M&S) approaches create a model of a real-world system that is linked to data sources and is used to simulate and predict the behavior of its real-world counterpart. On the face of it Digital Twins and M&S appear to be similar, if not the same. Is this actually the case? Are the two fields really separate or is Digital Twin research re-inventing the “M&S wheel”? To investigate these relationships, in this panel we will explore some contemporary innovations with Digital Twins and discuss whether or not Digital Twins is a contemporary “refresh” or “rebranding” of M&S or if there are exciting new synergies.

ABSTRACTThe purpose of this panel is to discuss the state of the art in digital twin for manufacturing research and practice from the perspective of the simulation community. The panelists come from the US, Europe, and Asia representing academia, industry, and government. This paper begins with a short introduction to digital twins and then each panelist provides preliminary thoughts on concept, definitions, challenges, implementations, relevant standard activities, and future directions. Two panelists also report their digital twin projects and lessons learned. The panelists may have different viewpoints and may not totally agree with each other on some of the arguments, but the intention of the panel is not to unify researchers' thinking, but to list the research questions, initiate a deeper discussion, and try to help researchers in the simulation community with their future study topics on digital twins for manufacturing.

This panel will discuss the inherent conflict between the application of (Discrete-Event) Simulation and Scheduling techniques to manage and optimise capacity and material flow in Semiconductor Frontend Manufacturing (wafer fabrication). Representatives from both industry and academia will describe advantages and shortcomings of the respective techniques, with a specific focus on challenges arising from the recent and anticipated future evolution of the nature of such manufacturing environments, and suggest solution approaches as well as research issues that need to be addressed.

Technology and product development have high priority in an advanced semiconductor manufacturing facility such as the Infineon Dresden fab. From the perspective of line performance this means that short cycle times for development lots have to be guaranteed to enable the required learning cycles. Long-term simulation is used in dynamic capacity planning to find a compromise between short cycle times for the development corridor and high utilisation of the installed tool capacity. All products in the fab run with customer-specific due dates. As such, negative side-effects caused by the accelerated development lot corridor through increased dispatch priorities have to be minimised. In turn, for day-to-day operations short-term simulation is used for early detection of bottleneck situations and other sudden resource availability problems. With focus on the development corridor, a Lot Cycle Time Forecaster was realised. The aforementioned manifold applications of discrete-event simulation are described in this paper in more detail.

Discrete Optimization via Simulation (DOvS) has drawn considerable attention from both simulation researchers and industry practitioners, due to its wide application and significant effects. In fact, DOvS usually implies the need to solve large-scale problems, making the efficiency a key factor when designing search algorithms. In this research work, MO-COMPASS (Multi-Objective Convergent Optimization via Most-Promising-Area Stochastic Search) is developed, as an extension of the single-objective COMPASS, for solving DOvS problems with two or more objectives by taking into consideration the Pareto optimality and the probability of correct selection. The algorithm is proven to be locally convergent, and numerical experiments have been carried out to show its ability to achieve high convergence rate.

Material flow forecast based on Short-Term Simulation has been established as a decision support solution for fine-tuning of Preventive Maintenance (PM) timing at Infineon Dresden. To ensure stable forecast quality for effective PM decision making, the typical tool uptime behavior needs to be portrayed accurately. In this paper, we present a hybrid tool down modeling approach that selectively combines deterministic and random down time modeling based on historical tool uptime behavior. The method allowed to approximate the daily uptime of reality in simulation. A generic framework to model historical down behavior of any distribution type, described by the two parameters Mean Time to Failure (MTTF) and Mean Time to Repair (MTTR) is also discussed.

Semiconductor wafer fabrication is one of the most challenging manufacturing environments around because of its complexity and large number of production steps involved as well as its capital intensity and the competitiveness of the semiconductor industry. This paper explains why - as opposed to a scheduling paradigm - simulation-based WIP management in conjunction with real-time dispatching is the more appropriate future approach for operations management in a semiconductor wafer fabrication facility.

Short Term Simulation (STS) that provides daily forecasts of work center performance has been deployed in Infineon Technologies for operational decision makings. To ensure good forecast accuracy, the STS requires high modeling fidelity, requiring good basic data quality for model building. Forecast accuracy is maintained through an Automatic Model Verification (AMV) engine. The AMV monitors and verifies discrepancies between simulation and reality for modeling elements such as process dedication, uptime, process time/throughput, sampling rate, and batch/stream size. It reports the verification results with a multi-layered view, at different levels of abstraction, and the gaps between simulation and reality are highlighted. The user can quickly identify gaps and make correction to the errors. In this paper, we give an insight to the complete workflow on how AMV helps to detect data issues, the options to resolve such issues and the positive effect to the simulation forecast quality.

Discrete event simulation (DES) has been established as a frequently used decision-support method in semiconductor manufacturing. One of the key application areas is the planning and scheduling of extended (several days) maintenance activities. The first stage of maintenance activity planning is conducted with a transient long-term simulation model with the focus on evaluating the effect of maintenance activity on the expected fab performance. Decisions such as wafer start reduction or adjustment of delivery commitments among affected work centers are made. The second stage of the planning is initiated several days before the start of the maintenance activities, where resource planning and scheduling of the activity is done through assessment of the expected WIP situation forecasted by a high fidelity online simulation model. In this paper, we will explain this simulation-based multi-stage approach for maintenance activity scheduling. The associated benefits and challenges will be presented with an example use case.

Aviation spare parts provisioning is a highly complex problem. Traditionally, provisioning has been carried out using a conventional Poisson-based approach where inventory quantities are calculated separately for each part number and demands from different operations bases are consolidated into one single location. In an environment with multiple operations bases, however, such simplifications can lead to situations in which spares -- although available at another airport -- first have to be shipped to the location where the demand actually arose, leading to flight delays and cancellations. In this paper we demonstrate how simulation-based optimisation can help with the multi-location inventory problem by quantifying synergy potential between locations and how total service lifecycle cost can be further reduced without increasing risk right away from the Initial Provisioning (IP) stage onwards by taking into account advanced logistics policies such as pro-active re-balancing of spares between stocking locations.

To make use of short-term simulation on an operational level, three aspects are essential. First, the simulation model needs to have a high level of detail to represent a small part of the wafer fab with sufficient precision. Second, the simulation model needs to be initialized very well with the current fab state. And third, the simulation results need to be available very fast, almost in real time. Unfortunately these conditions contradict each other. In particular, it takes a large amount of time to initialize a high precision full fab simulation model because of the huge amount of data. In this paper, we present the prototype of a fab driven simulation approach to overcome these time consuming limitations. We will show how it is possible to start a short-term simulation from the current fab state immediately, i.e., without further delay.

A cluster tool basically consists of one or more loadlocks where wafers enter and exit the cluster tool, two or more processing chambers where the processes are performed and one or more handlers that transport the wafers between loadlocks and processing chambers. This paper focuses on understanding the behaviour of a cluster tool system by simulation modelling, experimenting on the model and finding out the factors that influence the tool overall cycle time. The number of deposition chambers, number of robot grippers, clean cycle, MTTF and MTTR are identified as the most important factors for cycle time reduction at cluster tool in semiconductor wafer fabrication.