Optimizing semiconductor process recipe settings using hybrid meta-learning and metaheuristic approaches

Abstract

Achieving high-yield production in semiconductor thin film chemical vapor deposition (CVD) processes requires precise parameter settings, which often rely on costly R&D efforts and expert experience. This work proposes a novel approach combining meta-learning with metaheuristic algorithms to optimize these parameters more efficiently, particularly when experimental data is limited. Collaborated with a semiconductor manufacturer specializing in low-volume, high-variety products, we addressed the challenge of limited experimental data. To optimize new product processes, the company collected limited real experiment data. However, conventional soft computing methods often require extensive data for accurate prediction models, increasing computational time and costs. Therefore, we propose hybrid meta-learning and metaheuristic approaches to efficiently determine parameter settings. Meta-learning leverages historical data from similar tasks to train a neural acquisition function within the meta Bayesian optimization (MetaBO) framework. Two enhancements are proposed: employing the Halton sequence to reduce computational complexity and integrating three metaheuristic algorithms (genetic algorithm, particle swarm optimization PSO, and artificial fish school algorithm) to refine evaluation points and improve model quality. Experiments on benchmarking black-box functions and real semiconductor CVD processes show superior performance of our approaches over legacy MetaBO and other acquisition function methods, with the PSO-incorporated hybrid approach performing best.