Robust surrogate modeling for glass forming process

A critical challenge in manufacturing process optimization is that simulation-based surrogate predictive models often fail when confronted with real-world measurement uncertainties. In this study, we present a robust surrogate modeling approach applicable to simulation-based manufacturing process optimization, while accounting for real-world measurement uncertainties. Our methodological framework comprises the following: (1) generation of high-fidelity simulation data from finite element modeling, (2) training of diverse machine learning algorithms as surrogate models, (3) implementation of multi-objective hyperparameter optimization (MOHPO) to simultaneously optimize prediction accuracy and robustness against perturbations, and (4) quantification of real-world robustness through Monte Carlo sampling under simulated measurement uncertainties. Unlike previous predictive methods that focus solely on tuning prediction accuracy or incorporate robustness through model-specific techniques, our methodology systematically balances both objectives in a model-agnostic manner, requiring only simulation data for training. Using glass forming as a case study, we quantitatively evaluate six machine learning algorithms under temperature measurement uncertainties of ±3°C. In our experiments, multi-layer perceptrons achieve the best overall performance with mean squared error of nodal deviation < 0.2 while maintaining high robustness 0.6. Our approach generates a diverse set of Pareto-optimal solutions that allows post-training-and-optimization selection of the ideal model based on specific manufacturing requirements, eliminating the need to predefine the exact balance between accuracy and robustness before model development. This work represents a significant advancement in bridging the gap between idealized simulations and practical industrial applications by systematically accounting for measurement uncertainties in a model-agnostic manner.