Leveraging Bonding-Sensitive Descriptors with Machine Learning for CO₂​ Hydrogenation Barrier Prediction

Accurate prediction of activation energies in complex catalytic reaction networks is vital for rational catalyst design. Traditional Brønsted-Evans-Polanyi (BEP) relationships often provide insufficient accuracy for this purpose. We introduce a machine learning (ML) approach to predict activation energies for the CO₂​ hydrogenation reaction network, leading to C₁ (CO, methanol, formaldehyde, formic acid) and C₂ (ethanol, acetaldehyde, acetic acid) oxygenates on Pd surfaces. Our models are built upon a dataset of 312 elementary steps on Pd(111) and Pd(211), utilizing bonding-sensitive descriptors that effectively capture changes in reactant bonding during the reaction. Various ML regressors, including Random Forest, XGBoost, Support Vector, Kernel Ridge, Gaussian Process, and Deep Neural Networks (DNN), were evaluated, all consistently outperforming linear BEP scaling. The DNN achieved the highest accuracy, yielding a mean absolute error of 0.16 eV on a held-out test set. This methodology offers a robust and accurate alternative for predicting reaction barriers, with ongoing work extending its validation to ethylene formation pathways.