Neural Networks for Surrogate-assisted Evolutionary Optimization of Chemical Processes

In the chemical industry commercial process simulators are widely used for process design due to their extensive library of models of plant equipment and thermodynamic properties and their ease of use. Most of these simulators compute the steady-states of complex flowsheets, but their models are inaccessible and derivatives with respect to their model parameters are not available. Evolutionary algorithms are a suitable approach for the global optimization of such black-box models, but they require the evaluation of many individuals. Applications to industrial-size case-studies suffer from high computational times where the numerical simulations consume the majority of the time. This contribution proposes the use of neural networks as surrogate models to guide the evolutionary search. These models are trained multiple times during the evolutionary search and are used to exclude nonpromising individuals and to generate candidate solutions. We demonstrate the performance improvement due to the use of the surrogate models for a medium-size case-study of a chemical plant consisting of a reactor, a liquid-liquid separation and a distillation column. The results show that the required number of simulations can be reduced by 50%.