2025 AIChE Annual Meeting

Learning Transferable Zeonet Representations Using 3D U-Nets

Zeolites are remarkable candidates for use in separations processes because they can selectively adsorb molecules onto their nanopores. Roughly 250 zeolites have been synthesized, and over 300,000 have been predicted to exist within the bounds of thermodynamic feasibility. It is efficient and valuable to be able to identify the most promising structures before investing the time and effort to synthesize and test any of them. Molecular simulations have traditionally been used for this purpose, and although reliable, they incur considerable computational costs. We have explored deep learning as a way of bypassing these costs while preserving a competitive level of predictive power. In our recent work, we have shown that Zeonet, a 3D distance grid–based convolutional neural network (CNN), outperforms point cloud–, graph-, and vision transformer–based models at predicting Henry adsorption constants of long-chain hydrocarbons at the surfaces of various zeolites. These 3D CNNs require at least the order of 10,000 training samples to reach peak performance, and depending on the learning task, procuring and processing such large amounts of data can be infeasible. Using transfer learning, a pre-trained 3D CNN can be adapted to a new adsorption task and a similar performance level with only the order of 1,000 training samples. Here, we investigate the potential of 3D U-Nets in producing transferable data by training one to map distance grids to granular potential energy grids. This poster aims to communicate our design choices and preliminary data. The U-Net consists of down-sampling, encoding layers followed by up-sampling, decoding layers, with residual connections between corresponding encoder and decoder layers. Our current work involves comparing model performance after being trained on different settings of a custom loss function inspired by the Boltzmann distribution. We plan to evaluate the ability of the pre-trained encoder to transfer to various downstream learning tasks.