Metal–organic frameworks (MOFs) are promising materials for gas storage and separation because their pore environments can be chemically tuned across a vast design space. However, identifying the local adsorption environments that control performance remains challenging. Adsorbate probability distributions (APDs), obtained from molecular simulation, provide a direct atomistic picture of where adsorbates reside within MOF pores, with local maxima corresponding to binding sites. Despite their value, converging APDs using grand canonical Monte Carlo (GCMC) simulations is computationally expensive and therefore difficult to deploy in high-throughput materials discovery.

In this work, we present DeepAPD, an equivariant graph neural network for rapid prediction of APDs in MOFs. The model represents the framework as a periodic atomistic graph and uses spatial probe nodes to infer adsorbate probability values at arbitrary positions in the unit cell, enabling resolution-independent prediction of three-dimensional APDs. As a proof of concept, DeepAPD was trained on GCMC-derived APDs for approximately 23,000 MOFs for methane at 1 and 65 bar and xenon at 1 bar. Novel binding site analysis/extraction algorithms are employed to assess the performance of DeepAPD. The resulting models reproduce simulated APDs with high similarity and reliably identify high occupancy binding sites (positional errors of 0.15 -- 0.30 \AA). DeepAPD also outperforms simple guest–host energy-grid approaches, particularly under conditions where guest–guest interactions play a significant role in the adsorption free-energy landscape.

By replacing expensive GCMC sampling with a reliable surrogate model, DeepAPD enables APDs and binding sites to be generated in seconds to minutes rather than hours to days. The advantage of DeepAPD is expected to become even more advantageous for more complex adsorbates (e.g., \ce{CO2}, \ce{H2O}), where conventional GCMC sampling becomes more computationally demanding. This work introduces a route to incorporating atomistically resolved adsorption information into high-throughput MOF screening, binding-site mining, and future generative design workflows for porous materials.