Machine learning-enhanced analysis of catalyst particle size effects and performance prediction of platinum on carbon electrocatalysts

Understanding and optimizing the impact of catalyst particle size is critical for enhancing the performance of platinum-on-carbon (Pt/C) electrocatalysts in hydrogen evolution reactions (HER). However, conventional methods often fall short in capturing the complex, nonlinear, and multivariate relationships that govern particle size and catalytic efficiency. To address this, we present a novel machine learning-enhanced framework that combines CatBoost for granular analysis of catalyst particle size effects and XGBoost for high-accuracy performance prediction of Pt/C electrocatalysts. This dual-model integration offers a methodological advancement by jointly optimizing predictive performance and interpretability. The study reveals a linear relationship between the particle size of metal-based catalysts and the measured overpotential. Catalytic performance improves monotonically with decreasing particle size, particularly in the 1.0 to 5.0 nm range. Compared to the baseline Gradient Boosting Regressor (GBR), CatBoost improves the R² from 0.940 to 0.968 (ΔR²=+0.029). It also reduces mean absolute error (MAE) from 20.39 to 15.36 (ΔMAE=−5.03) and lowers mean squared error (MSE) from 1437 to 753.6 (ΔMSE=−683.4). Five-fold cross-validation confirms model consistency, with a mean R² of 0.941 (±0.014). Likewise, XGBoost enhances R² from 0.970 to 0.981 (ΔR²=+0.011), reduces MAE from 13.74 to 10.84 (ΔMAE=−2.9), and decreases MSE from 338 to 267.7 (ΔMSE=−70.3), supported by a cross-validated mean R² of 0.960 (±0.0159), validating model generalizability. XGBoost demonstrates robust performance in acidic electrolytes compared to the GBR model. Finally, interpretability analyses using SHAP and DALEX offer transparent insights into feature contributions to catalytic performance.