Multi-task optimization with Bayesian neural network surrogates for parameter estimation of a simulation model

Abstract

We propose a novel framework for efficient parameter estimation in simulation models, formulated as an optimization problem that minimizes the discrepancy between physical system observations and simulation model outputs. Our framework, called multi-task optimization with Bayesian neural network surrogates (MOBS), is designed for scenarios that require the simultaneous estimation of multiple sets of parameters, each set corresponding to a distinct set of observations, while also enabling fast parameter estimation essential for real-time process monitoring and control. MOBS integrates a heuristic search algorithm, utilizing a single-layer Bayesian neural network surrogate model trained on an initial simulation dataset. This surrogate model is shared across multiple tasks to select and evaluate candidate parameter values, facilitating efficient multi-task optimization. We provide a closed-form parameter screening rule and demonstrate that the expected number of simulation runs converges to a user-specified threshold. Our framework was applied to a numerical example and a semiconductor manufacturing case study, significantly reducing computational costs while achieving accurate parameter estimation.