Prediction of Hydration energies of Adsorbates at Pt(111) and Liquid Water Interfaces using Machine Learning

Aqueous-phase heterogeneous catalysis has many applications, including biomass reforming, Fischer-Tropsch synthesis, and electrocatalysis. Developing accurate models for these systems is essential for gaining mechanistic understanding and making predictions of activity and selectivity under reaction conditions. However, molecular modeling of solid-liquid interfaces is computationally demanding. To address this, we carried out machine learning analysis on an existing dataset comprising energies and free energies of solvation for 90 adsorbates on a Pt(111) surface. These adsorbates include intermediates from the decomposition of methane, methanol, ethylene glycol, and glycerol. We investigated the structure-property relationship by combining molecular descriptors with machine learning models. Eight machine learning approaches were compared. In general, machine learning models outperform molecular dynamics for computing the same properties and achieve RMSE < 0.1 eV for predicting the energies and free energies of solvation, which is within the standard error within the original dataset. R2 for energies of solvation are in general above our threshold value of 0.8 but only 0.72 for free energies of solvation. To achieve better regression for free energies of solvation, the dataset should be expanded. However, our machine learning model still outperforms molecular dynamics for predicting free energies of solvation. Feature importance analysis shows that while hydrogen bonds between water and the adsorbates contribute most strongly to machine learning model performance, a combination of different types of features is important to achieve strong predictive performance.