How Domain Knowledge can Improve Machine Learning Surrogates for Manufacturing Process Optimization – a Comparative Study

Béla H. Böhnke, Aleksandr Eismont, Clemens Zimmerling, Luise Kärger, Klemens Böhm

Abstract: In various industries, optimizing manufacturing parameters is vital for the efficient production of high-quality products. Traditional methods involve costly production trials and process tuning – particularly when dealing with complex processes and materials such as composites. High-fidelity simulations offer a cost-effective alternative. However, they can be computationally intensive, which often renders them impracticable for iterative optimization. Surrogate model-based optimization (SuMO) provides a solution by using efficient, data-driven approximations. However, existing approaches often overlook valuable domain knowledge, such as material behavior, spatial relationships and optimization objective. We investigate different types of knowledge varying in complexity, difficulty to incorporate and transferability to other domains. In numerical studies on composite manufacturing – specifically, textile draping – we demonstrate that integrating such domain knowledge improves prediction accuracy, reduces optimization iterations, and enhances overall outcomes.