Optimizing operations in semiconductor manufacturing facilities is challenging. The production line can be modeled as an NP-hard constrained flexible job-shop scheduling problem, intractable with mathematical optimization due to its scale. Therefore, decision-making in factories is dominated by handcrafted heuristics. Though machine learning-based approaches proved efficient in solving such problems, their applications are limited due to the lack of trust in the underlying black-box models, and issues with scalability for larger instances. This work presents a genetic programming-based method to generate explainable, improved dispatching heuristics. Our method outputs a set of human-readable dispatching strategies, verifiable by scheduling experts before deployment. In case of minor changes in the environment or the optimization objectives, the continued evolution of the candidate solutions is possible without starting the training process from scratch. The introduced method is evaluated on a simulator executing real-world scale instances. The resulting heuristics improve the key performance indicators of the generated schedules. Furthermore, the generated dispatchers are easy to integrate into existing industrial systems. These favorable properties make the method applicable to various large-scale, dynamic, practical scheduling scenarios, where adaptions to different environments go along with modest human effort limited to the design of a fitness function.

The goal of this work is to develop novel methods to solve the semiconductor fab scheduling problem. The problem can be modeled as a flexible job-shop with large instances and specific constraints related to special machine and job characteristics. To investigate the problem, we develop a tool to simulate small to large-scale instances of the problem. Using the simulator, we aim to develop new dispatching strategies using genetic programming and reinforcement learning.

Efficient production scheduling is an important application area of constraint-based optimization techniques. Problem domains like flow- and job-shop scheduling have been extensive study targets, and solving approaches range from complete and local search to machine learning methods. In this paper, we devise and compare constraint-based optimization techniques for scheduling specialized manufacturing processes in the build-to-print business. The goal is to allocate production equipment such that customer orders are completed in time as good as possible, while respecting machine capacities and minimizing extra shifts required to resolve bottlenecks. To this end, we furnish several approaches for scheduling pending production tasks to one or more workdays for performing them. First, we propose a greedy custom algorithm that allows for quickly screening the effects of altering resource demands and availabilities. Moreover, we take advantage of such greedy solutions to parameterize and warm-start the optimization performed by integer linear programming (ILP) and constraint programming (CP) solvers on corresponding problem formulations. Our empirical evaluation is based on production data by Kostwein Holding GmbH, a worldwide supplier in the build-to-print business, and thus demonstrates the industrial applicability of our scheduling methods. We also present a user-friendly web interface for feeding the underlying solvers with customer order and equipment data, graphically displaying computed schedules, and facilitating the investigation of changed resource demands and availabilities, e.g., due to updating orders or including extra shifts.