@Article{D3TA01911J,
author ="Liu, Yachan and Perez, Gustavo and Cheng, Zezhou and Sun, Aaron and Hoover, Samuel and Fan, Wei and Maji, Subhransu and Bai, Peng",
title  ="ZeoNet: 3D convolutional neural networks for predicting adsorption in nanoporous zeolites",
journal  ="J. Mater. Chem. A",
year  ="2023",
pages  ="-",
publisher  ="The Royal Society of Chemistry",
doi  ="10.1039/D3TA01911J",
url  ="http://dx.doi.org/10.1039/D3TA01911J",
abstract  ="Zeolites are one of the most widely used materials in the chemical industry due to their nanometer-sized pores that can adsorb and react upon molecules selectively. With hundreds of known framework topologies and hundreds of thousands of computationally predicted structures{,} the ability to rapidly predict zeolite performance allows researchers to prioritize their efforts on the most promising structures for a given application. Although the accuracy of forcefield-based atomistic simulations has advanced significantly in the past two decades{,} these simulations can be computationally expensive{,} especially for long-chain{,} complex molecules. We present ZeoNet{,} a representation learning framework using convolutional neural networks (ConvNets) and 3D volumetric representations for predicting adsorption in zeolites. ZeoNet was trained on the task of predicting Henry’s constants for adsorption{,} kH{,} of n-octadecane in more than 330{,}000 known and predicted zeolite materials. Employing a 3D grid based on the distances to solvent-accessible surfaces{,} a volumetric representation that can be generated efficiently{,} the best-performing ZeoNet achieved a correlation coefficient r2 = 0.977 and a mean-squared error MSE=3.8 in ln kH{,} which corresponds to an error of 9.3 kJ/mol in adsorption free energy. In comparison{,} a model based on hand-designed geometric features has values of r2 = 0.777 and MSE=36.6. ZeoNet is also relatively efficient and can process ≈ 8 structures per second on an Nvidia RTX 2080TI GPU{,} orders of magnitude faster than forcefield-based simulations. A systematic analysis was conducted to investigate how the choice of ConvNet architectures{,} the linear dimension (L) and spatial resolution (∆d) of the distance grids{,} batch size{,} optimizer{,} and learning rate impact the model performance. We found that ConvNets based on the ResNet architecture offer the best tradeoff between expressiveness and efficiency. The performance for all models reaches a plateau at L = 30−45 A and depends less sensitively on grid resolution{,} with a small benefit around ∆d = 0.30−0.45 ̊A. Finally{,} saliency maps were visualized to identify which regions of the materials contributed the most to model predictions. It was found{,} interestingly{,} that the predictions are driven primarily by the accessible pore volume rather than the region occupied by the framework atoms."}