Quantifying the Global Impact of Constraint Programming-Based Local Scheduling in Semiconductor Manufacturing

Abstract

The efficiency of semiconductor frontend manufacturing highly depends on the optimization of resource allocation. In academic works, scheduling methods, i.e., based on Constraint Programming (CP) or Mixed Integer Programming (MIP), are popular tools for solving this optimization problem. As the problem is NP-hard, complete optimization methods do not scale to the problem size and complexity required for realistic fab-wide scheduling. Therefore, the problem is often decomposed and solved locally. While we can track the solution gap or even compute exact solutions for sufficiently small local problem instances, the fab-wide impact of decomposition strategies is yet to be determined in a realistic setting. In this work, we empirically quantify the impact of locally applied optimization on the fab-wide throughput and tardiness. For this evaluation, we use a local CP-based optimization method to compute large-scale schedules in the dynamic and stochastic environment of a semiconductor frontend manufacturing facility. We investigate the local and global impact of our method in different settings using a high-fidelity simulation environment. The results show that the CP-based scheduling method is able to significantly improve locally over advanced expert-drafted heuristics. On the global scale, we show that local optimization delivers inconsistent and varying results, stressing the need for an injection of global information into the local problem formulation.