A stopping rule for completing computed phase diagrams with generative models

Phase diagrams constructed using known phases misjudge stability when key phases are missing. Generative models supply as many structures as needed, but how many structures to generate and which diagrams will change are unknown. We add 10,000 MatterGen structures per system, relaxed with a machine learning interatomic potential, to Materials Project (MP) data for six Li–transition-metal–O ternaries and track the hull convergence. The responses fall into two groups: Li–Ru–O, Li–Hf–O, and Li–Zr–O (70–107 MP structures) gain 4–9 stable candidates and exhibit total convex-hull shifts of 59.9–62.9 meV/atom, whereas Li–Mn–O, Li–Co–O, and Li–Ni–O (240–609) exhibit a minimal hull shift. When reduced to 70 MP structures, the latter gain a similar number of stable candidates by rediscovering removed hull phases. The hulls converge within 5,000–8,000 structures. After generating 8,000 structures, a stopping rule, tested using Li–Ir–O, decides whether more are needed. Density functional theory places seven Li–Ru–O compositions 5–83 meV/atom below the MP-only hull but rejects the potential’s false positives in well-explored systems. This gives a stopping rule and early diagnostic for incomplete phase diagrams.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-17
DOI
https://doi.org/10.26434/chemrxiv.15012360/v1
Primary Topic
Machine Learning in Materials Science
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A stopping rule for completing computed phase diagrams with generative models

Juo Kim, Kyoungmin Min, 박재휘, Yerim Jung et al.
3 links
ChemRxiv
Machine Learning in Materials Science
preprint

A stopping rule for completing computed phase diagrams with generative models

Juo Kim, Kyoungmin Min, 박재휘, Yerim Jung, Seungbae Kim
preprint en

Abstract

Phase diagrams constructed using known phases misjudge stability when key phases are missing. Generative models supply as many structures as needed, but how many structures to generate and which diagrams will change are unknown. We add 10,000 MatterGen structures per system, relaxed with a machine learning interatomic potential, to Materials Project (MP) data for six Li–transition-metal–O ternaries and track the hull convergence. The responses fall into two groups: Li–Ru–O, Li–Hf–O, and Li–Zr–O (70–107 MP structures) gain 4–9 stable candidates and exhibit total convex-hull shifts of 59.9–62.9 meV/atom, whereas Li–Mn–O, Li–Co–O, and Li–Ni–O (240–609) exhibit a minimal hull shift. When reduced to 70 MP structures, the latter gain a similar number of stable candidates by rediscovering removed hull phases. The hulls converge within 5,000–8,000 structures. After generating 8,000 structures, a stopping rule, tested using Li–Ir–O, decides whether more are needed. Density functional theory places seven Li–Ru–O compositions 5–83 meV/atom below the MP-only hull but rejects the potential’s false positives in well-explored systems. This gives a stopping rule and early diagnostic for incomplete phase diagrams.

ChemRxiv
Yonsei University (KR)
Yonsei University, Ministry of Trade, Industry and Energy, Korea Institute for Advancement of Technology, National Research Foundation of Korea, Ministry of Science and ICT, South Korea
Machine Learning in Materials Science
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon