A single reference-energy outlier nearly eliminates Co–O chemistry from a generative model's training distribution
- 1School of Materials Science and Engineering
- 2Kumoh National Institute of Technology
Abstract
An element-resolved audit of 44,688 oxide structures generated by MatterGen under chemical-system conditioning revealed that the aggregate composition compliance (73.27% exact match) concealed a severe chemistry-specific failure. Cobalt was omitted from 71.27% of the structures requesting it, against 8.12 ± 3.28% for the other 19 requested elements, and 98.17% of the cobalt omissions removed cobalt alone. Cell size, cation substitutability, stability gain and a general training-frequency effect did not account for the anomaly. Tracing the failure upstream showed that only one of the 8,924 cobalt oxides in the public Alex-MP-20 reference set passed the stability filter. The deficit was traced to a single anomalous reference-energy entry distorting the Co–O convex hull; excluding it raises the number of training-eligible cobalt oxides from 1 to 2,499. Four independent DFT calculations and the current Materials Project value agree within 0.024 eV atom-1, whereas the value implied by the released file deviates from them by 6.26 eV atom-1. A pre-registered intervention experiment showed that neither removal of the stability condition nor stronger diffusion guidance normalised the blind spot. These results show how a single high-leverage reference-energy anomaly can be amplified through stability filtering into a training deficit at the scale of a chemical family. Two audits that detect this class of failure without retraining are presented, namely an element-resolved condition-fidelity audit and a reference-hull vertex audit.
Related
- Built in Minutes, Checked by Script: Instructor-Made Interactive Chemistry Study Guides with Generative AI
- Escape Rooms, Arcades, and a Grain of Salt: Five Original Games for Introductory Chemistry, Built by One Instructor with Generative AI
- Robust generative transition-state models for unseen chemistry
- Integrating Generative AI and Data Science into General Chemistry Laboratories
- Chemistry, Generative AI, and English: implications for chemistry researchers from diverse language backgrounds
BibTeX
@misc{song2026single,
title = {{A single reference-energy outlier nearly eliminates Co–O chemistry from a generative model's training distribution}},
author = {Dowon Song},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008501/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008501/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - A single reference-energy outlier nearly eliminates Co–O chemistry from a generative model's training distribution
AU - Dowon Song
PY - 2026
DA - 2026/09/08
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008501/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008501/v1
AB - An element-resolved audit of 44,688 oxide structures generated by MatterGen under chemical-system conditioning revealed that the aggregate composition compliance (73.27% exact match) concealed a severe chemistry-specific failure. Cobalt was omitted from 71.27% of the structures requesting it, against 8.12 ± 3.28% for the other 19 requested elements, and 98.17% of the cobalt omissions removed cobalt alone. Cell size, cation substitutability, stability gain and a general training-frequency effect did not account for the anomaly. Tracing the failure upstream showed that only one of the 8,924 cobalt oxides in the public Alex-MP-20 reference set passed the stability filter. The deficit was traced to a single anomalous reference-energy entry distorting the Co–O convex hull; excluding it raises the number of training-eligible cobalt oxides from 1 to 2,499. Four independent DFT calculations and the current Materials Project value agree within 0.024 eV atom-1, whereas the value implied by the released file deviates from them by 6.26 eV atom-1. A pre-registered intervention experiment showed that neither removal of the stability condition nor stronger diffusion guidance normalised the blind spot. These results show how a single high-leverage reference-energy anomaly can be amplified through stability filtering into a training deficit at the scale of a chemical family. Two audits that detect this class of failure without retraining are presented, namely an element-resolved condition-fidelity audit and a reference-hull vertex audit.
ER -