(455k) Representation Learning for All-Silica Zeolites: Model Representations, Transfer Learning, and Multi-Task Learning

Zeolites are widely used in selective chemical separation and catalytic processes. Their molecular-level shape selectivity is determined by the detailed 3D arrangement of framework atoms. We have recently developed and optimized ZeoNet, a representation learning framework that employs 3D convolutional neural networks (ConvNets) and a volumetric representation based on distance grids.¹ ZeoNet was trained to predict Henry’s constants for n-octadecane adsorption in over 330,000 all-silica zeolites, achieving a correlation coefficient r² = 0.97 and a mean-squared error MSE of 3.8 in lnk_H. In this study, we compared ZeoNet with other common representations, including 2D multi-view distance images, 3D multi-channel grids, point clouds based on atom coordinates and on surface sampling, and various graph networks. We found that 3D volumetric grids significantly outperformed other representations and employing a 3D ResNet18 architecture with single-channel distance grids achieved the best balance between computational efficiency and accuracy. Transfer learning was performed to predict adsorption for five other hydrocarbons with limited data, demonstrating that a pre-trained ZeoNet model could be effectively transferred to predict adsorption for these additional hydrocarbons with only hundreds of training samples. Finally, we also explored a multi-task learning problem, comparing different approaches to utilize available adsorption data for all six adsorption tasks simultaneously with a focus on obtaining the most transferable representation.

1. Y. Liu, G. Perez, Z. Cheng, A. Sun, S. Hoover, W. Fan, S. Maji, and P. Bai, “ZeoNet: 3D Convolutional Neural Networks for Predicting Adsorption in Nanoporous Zeolites,” J. Mater. Chem. A 11, 17570-17580 (2023).