Predicting Adsorbate-Solvent Interaction Energies in Zeolite Pores Using Convolutional Neural Networks with Attention...

Solvent effects play a critical role in determining reaction thermodynamics and kinetics in liquid-phase heterogeneous catalysis, particularly in confined nanoporous environments such as zeolite pores. However, accurate prediction of solvation thermodynamics remains computationally prohibitive due to the need for extensive first-principles calculations on diverse solvent configurations. In this work, we present an attention-enhanced three-dimensional convolutional neural network (3D-CNN) framework to predict solvation interaction energies (ΔEint) of adsorbates in Ti-faujasite (Ti-FAU) zeolite pores. Molecular dynamics (MD) snapshots of C1-C3 oxygenates in aqueous and water-methanol solvent environments are transformed into multichannel voxel grid representations that encode spatially resolved atomistic information. The 3D-CNN architecture incorporates separate adsorbate and solvent branches and employs convolutional block attention modules to dynamically prioritize spatial regions and feature channels relevant to adsorbate-solvent interactions. Trained on density functional theory (DFT)-derived interaction energies from a multiscale modeling workflow, the model achieves a mean absolute error of 0.08 eV for individual configurations and 0.04 eV for ensemble-averaged solvation energies, approaching DFT accuracy while substantially reducing computational cost. Cross-validation (CV) using simulation-system, solvent-composition, and pore-type splits demonstrates predictive performance on held-out systems within the represented chemical domain and across the solvent and pore conditions considered in this study. Analysis of spatial attention maps and feature attributions reveals that the model emphasizes adsorbate-proximal solvent structures and hydrogen-bonding interactions, consistent with physical intuition. This framework enables efficient and interpretable prediction of solvation interaction energies in confined systems and provides a scalable pathway for mapping structural information sampled from force-field-MD to DFT targets.