Explainable artificial intelligence for chemical process informatics: Energy-efficient design of styrene monomer production

• Global optimization of styrene plant design with explainable AI. • Bayesian optimization efficiently identified low-energy production conditions. • SHapley Additive Explanations (SHAP) guided interpretation of the process model. • SHAP-based feature selection improved model generalization performance. Parameter optimization in process design for chemical products is ideally performed at the system-wide level. However, such optimization presents a multivariate, multi-objective problem, requiring an extensive number of simulations. Here we optimized the parameters of the styrene monomer production process using an adaptive experimental design based on Bayesian optimization. As a result, we efficiently explored low-energy-consumption regions and successfully identified energy-efficient design points with a reduced number of simulations. Additionally, we analyzed the machine learning model using the SHapley Additive exPlanation (SHAP) method to visualize exploration trends and provide phenomenological interpretations. To further explore applications of SHAP analysis, we conducted feature selection and parameter tuning based on SHAP values. Feature selection using SHAP values significantly improved the generalization performance of the model, while parameter adjustments based on SHAP values further reduced energy consumption from the initially identified optimal points. This study represents the first attempt to integrate global optimization of overall chemical process design parameters with explainable artificial intelligence. Nevertheless, we also identified several challenges, such as the balance between exploration and exploitation and the limitations of local representations in the machine learning model. The framework combining Bayesian optimization and SHAP-based analysis developed in this study is applicable to various chemical processes. Our technical scheme has the potential to advance research on machine learning-driven process design and contribute to its industrial implementation in the future.