ChemSPAR: An auditable, chemistry-constrained framework for sparsifying periodic graphs of metal–organic frameworks

Graph neural networks (GNNs) for metal–organic frameworks (MOFs) usually connect every atom pair within a cutoff, so most edges are not chemical bonds. We found no sparsification method that guarantees the survival of chemically essential contacts and records each removal decision. ChemSPAR prunes periodic contact instances of crystal graphs under a prespecified chemistry taxonomy and protection gate, accepts any edge-ordering policy, and logs every removal attempt. On 5000 QMOF structures we compared gravity, distance, random and Black-Hole-Strategy-derived orderings, each with and without the gate, structurally up to a removal fraction of 0.95 and, at 0.5, with a CGCNN-style model for the PBE bandgap (five paired seeds, pre-specified analysis; full-graph and 3.5 Å cutoff references). At 0.5 the gate kept every protected bond, whereas ungated random and BHS-derived ordering removed about half of them; for these bond-blind orderings, the paired gated-minus-ungated difference in test MAE was −0.0269 and −0.0428 eV (Holm-adjusted p ≤ 0.002 ; lower in every seed). For distance ordering the gate changed few structures at this fraction, and its accuracy came from the ordering itself: distance ordering under the gate was the most accurate gated policy, kept complete local shells within 3.0 Å for 99.5% of atoms, and was equivalent to the full graph within a pre-specified ± 0.02 eV margin (difference −0.0095 eV, 95% CI −0.0163 to −0.0026) while halving training time per epoch. A plain 3.5 Å cutoff was about as accurate and cheaper to build, but lost protected bonds in 25 structures and provides no decision record.