A catalyst-informatics framework for Ga-Pt supported liquid alloy in propane dehydrogenation is presented. The framework, combining statistical analysis and machine learning, identifies key factors influencing catalyst performance and highlights the need for larger datasets for predictive design.

Abstract

Propane dehydrogenation (PDH) is limited by rapid catalyst deactivation. Supported catalytically active liquid metal solutions (SCALMS) based on Ga–Pt alloys offer high selectivity and coke resistance, yet their vast compositional and operational design space hampers efficient optimization. We compiled and FAIR-formatted 198 PDH experiments on Ga-Pt SCALMS, distilling 149 complete cases with 20 descriptors covering synthesis, support, metal loadings, reaction conditions, and four key performance indicators: low deactivation, high selectivity, conversion, and productivity. Exploratory statistics revealed strong Ga–Pt loading covariance, pretreatment-temperature effects on stability, and a distinctive high-conversion/high-selectivity but fast-deactivating regime for Ga₂O₃-Pt catalysts prepared reductively on CARiACT silica. Principal-component analysis captured 34% of variance in two dimensions, isolating clusters linked to support and pretreatment protocols. Feature-reduced datasets fed three machine-learning regressors; extreme gradient boosting achieved the best extrapolation for productivity (R ² = 0.58), Random forests best predicted deactivation (R ² = 0.43), while support vector regression yielded the most accurate conversion predictions (R ² = 0.68). SHAP analysis ranked pretreatment temperature, Ga/Pt ratio, and time-on-stream as dominant drivers of KPI variance, aligning with SCALMS mechanistic expectations. Validation on six new experiments confirmed model fidelity within ± 15% for conversion, productivity, and deactivation. The combined statistical-predictive workflow constitutes a catalyst-informatics framework that guides catalyst development based on experimental data and highlights the need for larger, standardized datasets to reach truly predictive design of liquid–metal catalysts.