Machine Learning with Heuristic Search-based Hybrid Framework for Cycle Time Optimization in Semiconductor Production

This paper aims to optimise cycle time (CT) in semiconductor wafer production, a critical factor for enhancing operational efficiency and competitiveness in the semiconductor manufacturing industry. A hybrid methodology, based on statistical analysis and machine learning (ML) techniques, is developed to identify the optimal combination of key performance indicators (KPIs) for individual tools to minimise CT. To achieve this, hyperparameter tuning and model optimisation are performed using Sequential Quadratic Programming (SQP), Particle Swarm Optimization (PSO), and Genetic Algorithm (GA), with a focus on identifying the most effective optimisation technique. The optimisation process incorporates constraints on KPIs, introducing additional complexity and necessitating robust constraint-handling mechanisms. A hierarchical decomposition approach is employed to systematically address the problem, achieving significant reductions in production cycle time. The experimental study suggests that the random forest algorithm with GA significantly outperforms other techniques in terms of CT reduction.