The Kohn-Sham equations underlie many important applications such as the discovery of new catalysts. Recent machine learning work on catalyst modeling has primarily focused on prediction of the energy, but has so far failed to demonstrate widespread generalization to new structures with precision meeting target thresholds. Here we investigate another approach based on the pointwise learning of the Kohn-Sham charge-density. On a new dataset of bulk catalysts with charge densities, we give evidence of combinatorial generalization to 45