Surrogate Model Selection and Practical Guidelines for Constrained Bayesian Optimization in Chemical Process Design: Insights from Ammonia Synthesis Simulations

Constrained Bayesian optimization (BO) is a promising approach for optimizing chemical processes under limited evaluation budgets. In this study, BO was applied to an ammonia production process, and three surrogate models—Gaussian process regression (GPR), random forest (RF), and Bayesian linear regression (BLR)—were systematically compared. GPR achieved the highest optimization performance in terms of best-observed profit, whereas RF provided superior robustness with consistent identification of feasible solutions. BLR showed balanced performance with lower computational cost. Sensitivity analysis revealed that system behavior is governed by a small subset of variables, indicating low effective dimensionality. This structural property explains the strong performance of GPR. In addition, the relative performance of surrogate models was found to depend on economic conditions and optimization settings, as changes in dominant variables altered model suitability. Quantitatively, GPR achieved an average best-observed profit 10.4% and 5.2% higher than RF and BLR, respectively; RF identified 40.3% more feasible solutions than GPR; and BLR required only approximately 1/13 and 1/319 of the BO computation times required by GPR and RF, respectively. These results suggest that surrogate model selection in constrained BO should be adapted to problem structure and computational requirements, rather than relying on a single model.