Pareto Optimization of a Cobalt Carbide Catalyst for Syngas Conversion: Balancing Activity and Selectivity via Interpretable Active Learning
TL;DR
An active-learning (AL)-assisted machine-learning (ML) framework was developed to predict the CO conversion and light-hydrocarbon selectivity over cobalt carbide catalysts, and the predicted Pareto-optimal catalyst configurations showed good agreement with representative reported experimental catalysts.
Abstract
Cobalt carbide catalysts have attracted significant attention due to their high activity and light-hydrocarbon selectivity in the direct conversion of syngas, yet navigating their multidimensional parameter space typically relies on a trial-and-error screening process. In this work, an active-learning (AL)-assisted machine-learning (ML) framework was developed to predict the CO conversion and light-hydrocarbon selectivity over cobalt carbide catalysts. A data set containing 140 catalyst systems and 658 experimental records has been mined from experimental literature. To improve model generalization for small-sample data sets, catalyst-composition-stratified sampling was integrated with AL during model development. Several regression algorithms were systematically compared by using grid search and cross-validation. The optimal model achieved test set R2 values of 0.86 and 0.87 for CO conversion and light-hydrocarbon selectivity, respectively. Interpretation using Shapley additive explanations indicated that CO conversion is primarily influenced by reaction conditions, particularly temperature and H/C ratio, whereas light-hydrocarbon selectivity shows stronger dependence on catalyst structural features, including pore size and promoter content. Partial dependence plot analysis further revealed nonlinear interactions among key descriptors, suggesting different controlling factors for the catalytic activity and product distribution. Pareto optimization revealed an intrinsic tradeoff between CO conversion and light-hydrocarbon selectivity, and the predicted Pareto-optimal catalyst configurations showed good agreement with representative reported experimental catalysts. The proposed framework demonstrates the potential of integrating AL, interpretable ML, and Pareto optimization for catalyst evaluation and multiobjective optimization using small-sample catalytic data sets.