Gaussian Process Surrogates for Robust Optimization of Multi-Stage Manufacturing Processes

This doctoral thesis concerns the optimization of multi-stage manufacturing processes. In multi-stage manufacturing, the overall manufacturing process comprises several subprocesses, i.e. stages. Because real-world manufacturing processes are costly in terms of materials, labor hours, energy, and CO2 emissions, physics-based simulations can be used to represent individual manufacturing stages. One drawback of physics-based simulations is their computational complexity, therefore, the optimization approach studied in this thesis is based on machine learning surrogates of physics-based simulations. Researched optimization methods are based on Bayesian optimization (BO) with Gaussian process (GP) surrogates. Approaches proposed in this thesis concern the handling of epistemic surrogate model uncertainty and aleatoric manufacturing process uncertainty in BO. Optimization is considered towards a target, not minimization or maximization like in standard BO. Further, the interaction of process stages is considered in optimization, by optimizing on a multi-stage level, rather than individual optimization of single stages. Objectives do not rely solely on a final multi-stage output, but also consider interim stages to ensure lowest out-of-tolerance w.r.t. the overall manufacturing process. The methods researched in this thesis, propose fulfilling approaches for robust optimization of manufacturing processes. Use cases handled focus on hot metal forging production for aerospace.