Predicting hydrogen storage in MOFs: Representation matters

Metal–organic frameworks (MOFs) offer an expansive, tunable design space for hydrogen storage, but the sheer size of this space demands efficient screening methods. Machine learning can fill this role; however, a model can only be as good as the information it receives, raising the question of how the choice of structural representation affects predictive performance. To investigate this, we benchmark three representations under a unified protocol: namely, geometric descriptors, energy voxels, and molecular point clouds. Comparing predictive accuracy, data efficiency, and transferability to unseen MOF databases, we find that energy voxels, which encode host–guest interactions, consistently outperform geometry-only approaches, with the gap widening as local interactions become the dominant factor governing uptake. Our results highlight that the choice of representation, and the physical information it encodes, largely determines what a model can learn.