Computationally‐Efficient Environmental and Economic Multi‐Objective Optimization of a Methanol Production Process via Surrogate Modeling

Process simulation is a powerful tool in the Process Systems Engineering (PSE) field, in particular for optimization tasks. However, the computational times involved in these activities may become prohibitive for complex processes. As an alternative, data-driven strategies, such as surrogate models, have been widely adopted. Surrogate models are typically trained on data generated from specifically designed simulation runs. The computational efficiency of these designs has been addressed in the literature by minimizing the total number of simulations. However, the execution time of each simulation may be potentially reduced by shortening the transient period between consecutive simulations, for example, by minimizing the Euclidean distance between them. Sorting the simulations of the design to minimize the total traveled distance describes a typical Traveling Salesman Problem (TSP) scenario. This work analyzes the effect of four random and sorted one-shot experimental designs, composed of 50 samples, on the surrogate model training and surrogate-based optimization of a methanol synthesis process: DoE 1) Latin Hypercube (LHS), DoE 2) maxmin LHS, DoE 3) maxmin LHS sorted with nearest neighbors, and DoE 4) maxmin LHS sorted with 2-opt. Results showed that sorted DoEs improved the surrogate model accuracy by reducing its relative error by 0.3%. In addition, the overall computational time required diminished by around 14%. The most efficient experimental design was DoE 4, which was used to train a model later employed to optimize the OPEX and CO (Formula presented.) emissions of the methanol process, resulting in reductions of 15.0% and 11.4%, respectively.