Chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r2SCAN stability verdicts in the GNoME materials database
Abstract
GNoME's stability labels rest on a Materials Project-compatible GGA/GGA+U workflow. Its authors report that 84% of discovered binaries and ternaries retain a negative decomposition energy under r2SCAN and 86.8% of tested quaternaries remain stable on the r2SCAN hull. We resolve that aggregate by chemistry over the 139,541 GNoME materials released with r2SCAN recomputations, a subset the dataset authors selected rather than sampled, using cluster bootstrap intervals over 8,429 structural-prototype proxy clusters. The aggregate conceals a 26.7-point spread at fixed element count, from 4.0% [3.3, 4.7] for plain quaternary intermetallics (no lanthanide, no 3d metal) to 30.7% [27.7, 33.8] for lanthanide-free 3d-containing quaternary oxides, a pair identified post hoc against a prospectively fixed threshold. Under our reading of the release's "All" reference set the two cells sit at 14.5% and 37.4%, preserving a 22.9-point spread. Presence of a Materials Project Hubbard-U-list element is strongly associated with disagreement, conditional on the other covariates, in a model over all 139,541 materials (cluster-robust odds ratio 3.45 [3.13, 3.80]). Neither selective U application nor hull proximity explains the association away: disagreement stays elevated outside the U-treated regime, and a natural cubic spline in the continuous PBE decomposition energy moves the odds ratio in the 117,018-material model only from 3.53 to 3.43 [3.07, 3.83]. For allocating a verification budget, composition supplements hull proximity rather than replacing it: over held-out cluster splits at a 20% budget, the two together surface 45.8% of disagreements against 40.4% for hull distance alone. An exploratory comparison against previously catalogued materials changes direction between populations. We measure disagreement between two workflows, not accuracy.
Related
- Two-Dimensional Materials toward CMOS-Compatible Scalable Quantum Hardware
- Neutral atom quantum computing for materials science and quantum chemistry
- Culinary Chemistry as a Platform for Interfacial Materials Design
- Measurement-efficient energy estimator for quantum chemistry of photodynamic materials
- Nuclear-Electronic Orbital Theory: Exploring Nuclear Quantum Effects in Chemistry, Biology, and Materials.
BibTeX
@misc{aduru2026chemistryresolved,
title = {{Chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r2SCAN stability verdicts in the GNoME materials database}},
author = {Karthikeya Aduru},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.33101},
eprint = {2610.33101},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.33101},
}
RIS
TY - UNPB
TI - Chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r2SCAN stability verdicts in the GNoME materials database
AU - Karthikeya Aduru
PY - 2026
DA - 2026/10/22
PB - arXiv
DO - 10.48550/arxiv.2610.33101
UR - https://arxiv.org/abs/2610.33101
AB - GNoME's stability labels rest on a Materials Project-compatible GGA/GGA+U workflow. Its authors report that 84% of discovered binaries and ternaries retain a negative decomposition energy under r2SCAN and 86.8% of tested quaternaries remain stable on the r2SCAN hull. We resolve that aggregate by chemistry over the 139,541 GNoME materials released with r2SCAN recomputations, a subset the dataset authors selected rather than sampled, using cluster bootstrap intervals over 8,429 structural-prototype proxy clusters. The aggregate conceals a 26.7-point spread at fixed element count, from 4.0% [3.3, 4.7] for plain quaternary intermetallics (no lanthanide, no 3d metal) to 30.7% [27.7, 33.8] for lanthanide-free 3d-containing quaternary oxides, a pair identified post hoc against a prospectively fixed threshold. Under our reading of the release's "All" reference set the two cells sit at 14.5% and 37.4%, preserving a 22.9-point spread. Presence of a Materials Project Hubbard-U-list element is strongly associated with disagreement, conditional on the other covariates, in a model over all 139,541 materials (cluster-robust odds ratio 3.45 [3.13, 3.80]). Neither selective U application nor hull proximity explains the association away: disagreement stays elevated outside the U-treated regime, and a natural cubic spline in the continuous PBE decomposition energy moves the odds ratio in the 117,018-material model only from 3.53 to 3.43 [3.07, 3.83]. For allocating a verification budget, composition supplements hull proximity rather than replacing it: over held-out cluster splits at a 20% budget, the two together surface 45.8% of disagreements against 40.4% for hull distance alone. An exploratory comparison against previously catalogued materials changes direction between populations. We measure disagreement between two workflows, not accuracy.
ER -