Multi-Objective optimization of injection molding parameters: an experimental SHAP-Guided memetic NSGA-II approach

Abstract

Efficient parameter optimization in plastic injection molding remains challenging due to the conflicting goals of minimizing energy consumption and cycle time while maintaining target part-weight accuracy and controlling part-to-part consistency under unchanged process settings. Many existing approaches rely on simulation-based datasets that do not fully capture the behavior observed under real manufacturing conditions, limiting their practical transferability. This study proposes an experimentally grounded multi-objective optimization framework that integrates real production data, surrogate modeling, and a SHAP-guided memetic extension of NSGA-II. The methodology targets three key performance metrics: energy consumption, deviation from target part weight, and cycle time, while imposing a predicted within-batch weight-variability constraint under unchanged process settings. SHAP-derived feature attributions are used not only for model interpretation but also as problem-specific information to guide the local search stage. Three strategies are benchmarked: standard NSGA-II, NSGA-II with random local search, and the proposed SHAP-guided local search. Results show that explainability-guided search accelerates the early approach to high-quality Pareto regions and improves the use of function evaluations. Finally, representative Pareto-optimal configurations covering objective-oriented and compromise solutions are reproduced on the injection-molding machine, enabling direct comparison between predicted and experimentally measured responses and providing physical validation of the optimized operating conditions.