Answers are grounded in this paper and cite their sources.
This work audits the released Hessian Interatomic Potential checkpoints (hip_v3.ckpt, hip_v3_cf.ckpt, and hesspred_v1.ckpt) on the 710-molecule PubChem size-generalization benchmark (N = 30–100 atoms), where the mean Eckart-projected eigenvalue error of hip_v3.ckpt on the 702 valid 3D geometries reaches 1.8234 eV/(Ų·amu) (CPU), more than 300 times the in-distribution HORM value (0.00545). Most of this degradation is traced to three dataset and implementation artifacts: (1) eight reference files (PubChem_528*.h5) are flat 2D drawings (Δz = 0 Å, r_min = 0.074–0.620 Å) with a mean raw Eckart error of 385.48; (2) the 32-neighbor cap in torch_cluster.radius_graph truncates neighbors on the 5.0 Å backbone without sorting by distance, removing 649 directed covalent bond edges (r < 1.65 Å) in 227 of 543 high-degree molecules on CPU and 615 edges in 149 molecules under the Tesla T4 CUDA index rule; and (3) the default 100 Å Hessian cutoff passes atom pairs through radial weights that are sparsely trained across 7.5–14 Å and remain at initialization beyond 14.2 Å (accounting for about 13% of the error on larger molecules). Selecting the 32 nearest neighbors by distance (Near-32) and setting cutoff_hessian = 7.5 Å without retraining restores permutation equivariance to 1.4 × 10⁻⁶ eV/Å, matches the default graph on all structures with max_deg_5Å ≤ 32, reduces the 90th-percentile force MAE from 0.1410 to 0.0074 eV/Å, and lowers the mean Eckart error 139-fold to 0.0131 eV/(Ų·amu). A model-level Tesla T4 GPU evaluation on the 543 high-degree molecules confirms the same raw 100 Å Eckart baseline (1.868 vs. 1.877 on CPU) and a 4.6-fold larger error than Near-32 at 7.5 Å, while residual size-dependent error rises from 0.0103 (N = 30–39) to 0.0189 (N = 100). On a 10% stratified subset (n = 71), the same recipe lowers hip_v3_cf.ckpt Eckart MAE 548-fold (11.2422 → 0.0205) while leaving its scalar energy-gradient forces unimproved (0.1346 → 0.1306 eV/Å), whereas recovery on hesspred_v1.ckpt is incomplete (1.3845 → 0.2366 at its native k = 20 cap).
No takes yet. Share an insight, caveat, or question.
Abhijit Chatterjee (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: