Hybrid analytical surrogate-based process optimization via Bayesian symbolic regression

Modular chemical process simulators are widespread in chemical industries to design and optimize production processes with sufficient accuracy. However, convergence issues and entrapment in local optima during process optimization are still challenges to overcome. To circumvent them, surrogate models of first principles simulations have attracted attention as they are easier to handle, with hybrid surrogates combining data-driven surrogate models with mechanistic equations becoming particularly appealing. In this context, this work explores the use of Bayesian symbolic regression to construct and globally optimize hybrid analytical surrogate models of process flowsheets, where some units are approximated with tailored analytical expressions rather than with neural networks or Gaussian processes, which might be harder to globally optimize. Comparing with other prevalent black-box surrogate modeling & optimization approaches, such as kriging and Bayesian optimization, we find that our approach can find better solutions than those identified with pure black-box methodologies, yet model building is much more computationally demanding.