Machine Learning-Guided Catalyst Design for Ethane–CO₂ Co-Conversion

Here, 511 records from 59 publications were curated into 509 canonical entries, including 504 valid C2H4 yields. Six machine learning algorithms were evaluated using fold-contained preprocessing, DOI-grouped validation, chronological validation, and 50 repeated DOI-stratified holdouts.

Gradient Boosting and XGBoost showed the most consistent cross-literature performance, with mean repeated-holdout R2 values of 0.539 ± 0.115 and 0.544 ± 0.105, respectively. Convex stacking reduced variability but did not improve mean predictive accuracy. Subgroup analysis revealed weaker transfer to non-Cr catalysts.