Multiobjective Bayesian Optimization Framework for the Synthesis of Methanol from Syngas Using Interpretable Gaussian...

Area of Science:

• Catalysis and Reaction Engineering
• Renewable Energy Technologies
• Computational Chemistry and Materials Science

Background:

• Methanol is a key renewable fuel and hydrogen energy carrier, driving interest in efficient production methods.
• Synthesis gas (CO/H2) conversion to methanol involves two primary catalytic processes.
• Machine learning (ML) offers powerful tools for advancing reaction informatics and process optimization.

Purpose of the Study:

• To model carbon monoxide (CO) conversion and methanol selectivity using Gaussian process regression (GPR).
• To capture and quantify prediction uncertainty in methanol synthesis.
• To identify key operational factors influencing GPR model predictions.

Main Methods:

• Gaussian Process Regression (GPR) was applied to experimental data for CO conversion and methanol selectivity modeling.
• Shapley Additive exPlanations (SHAP) were used to interpret the GPR model and identify influential input features.
• Bayesian optimization was employed within a multiobjective framework to find optimal operating conditions.

Main Results:

• The GPR model demonstrated high accuracy in predicting CO conversion and methanol selectivity compared to other ML models.
• Inlet mole fraction of CO (Y(CO, in)) and net inlet flow rate (Fin(nL/min)) were identified as critical features influencing predictions.
• Interpretable GPR models facilitated the identification of optimal operating points for maximizing both conversion and selectivity.

Conclusions:

• GPR provides an accurate and interpretable approach for modeling methanol production processes.
• Understanding key input features enables targeted optimization for enhanced methanol synthesis.
• This work contributes to the development of efficient and sustainable methanol production technologies.