(451c) Developing 3D Convolutional Neural Networks (ZeoNet) for Predicting Adsorption in Nanoporous Zeolites

Zeolites are nanoporous materials widely applied in the chemical industry with many potential applications in chemical separations and catalysis. In this work, we present ZeoNet, a representation learning framework using 3D convolutional neural networks (ConvNets) and a new volumetric representation, to predict Henry’s constants for adsorption of long-chain hydrocarbon molecules in all-silica zeolites. The impact of various factors on model performance was systematically investigated and optimized, including ConvNet architecture, the linear dimension and spatial resolution of the volumetric grids, batch size, optimizer, and learning rate. The best-performing ZeoNet achieved a correlation coefficient r² = 0.977 and a mean-squared error MSE = 3.8 in lnkH, which corresponds to an error of only 9.3 kJ/mol in adsorption free energy. Additionally, feature attribution and visualization of trained models showed that the predictions by ConvNets are driven primarily by the accessible pore volume rather than the region occupied by framework atoms. Our study provides insights into how 3D ConvNets can best be used to represent extended materials systems as well as a collection of pretrained models for use in other applications where training samples are less abundant.