Machine learning (ML) models have been widely used as efficient surrogates for costly molecular simulations to predict gas adsorption in nanoporous materials for gas storage and separation applications. The “black box” nature of ML, however, remains a significant barrier between predictions and the design of novel nanoporous materials. In this work, we introduce PoroNet, a new graph neural network architecture built on a graph representation of the pore network (i.e., pore graph). In a pore graph, nodes represent individual pores and edges represent pore connections. PoroNet shows highly accurate predictions of gas adsorption capacity on benchmark datasets, which include the simulated adsorption data of spherical molecules (Kr and Xe) and linear alkane molecules (ethane and propane) in metal-organic frameworks (MOFs) under various pressures and temperatures. More importantly, pore-level contribution to the adsorption can be learned using PoroNet through both direct supervised learning and as an emergent property while fitting the total adsorption capacity. In the direct supervised learning experiments, we show that PoroNet is data-efficient in some cases, achieving comparable performance to the standard approach with only a fraction of simulation runs needed for model training. The pore-level contribution helps explain the ML predictions of the total adsorption behavior, identify the key pore properties that govern adsorption, and provide significant insights into pore engineering. We demonstrate that PoroNet is a powerful tool for high-throughput MOF screening and derivation of valuable design rules for hydrogen storage applications.
