Global optimization of symbolic surrogate process models based on Bayesian learning

In this work, we address the global optimization of process surrogates using Bayesian symbolic regression and deterministic global optimization algorithms. In contrast to other surrogates of process models that are hard to (globally) optimize, e.g., artificial neural networks or Gaussian processes, symbolic regression leads to a closed-form mathematical expression describing the observed data that can subsequently be globally optimized using off-the-shelf deterministic solvers. After providing an introductory example, we show the capabilities of our approach in the optimization of a methanol production plant. We further discuss the model accuracy, CPU times for model building and optimization, and outline the advantages and limitations of the proposed strategy.