Finite-temperature dynamic stability separates foundation machine-learning interatomic potentials that harmonic benchmarks rank equally
- 1Purdue University
Abstract
Foundation (universal) machine-learning interatomic potentials (MLIPs) are now used routinely as cheap stand-ins for DFT in high-throughput stability screening, but they are benchmarked almost exclusively against harmonic (0 K) phonons. Dynamic stability is physically a finite-temperature property, and an important class of materials, including cubic perovskites, bcc refractory metals and cubic fluorites, is harmonically unstable but thermally stabilised by anharmonicity. We test five foundation MLIPs (MACE-MP-0, CHGNet, ORB-v2, SevenNet-0, MatterSim) on this regime using a curated set of references whose finite-temperature behaviour is documented in the experimental and DFT literature, so that no new DFT ground truth is required. We introduce a cheap quantum selfconsistent-harmonic (SCHA) “soft-mode free energy” screen that evaluates every imaginary commensurate mode and calls the high-symmetry phase unstable if any of them condenses, and we cross-validate it against the gold-standard multi-mode stochastic SCHA (SSCHA). We report three findings together with a practical ensemble-disagreement guardrail. (i) At the harmonic level the models divide: MatterSim and SevenNet-0 reproduce every documented soft mode (accuracy 1.00), whereas MACE-MP-0 and CHGNet soften the bcc Zr/Hf instabilities to zero (15% false-stable rate) and ORB-v2 misses the SrTiO3 soft mode (8%). (ii) Harmonic accuracy does not predict finitetemperature accuracy in either direction: CHGNet is the worst model harmonically yet the second best at finite temperature, while MatterSim is harmonically perfect and only midtable, and on the matched set the two are non-significantly concordant at 100 K and significantly anti-associated at 300 K. Screening every imaginary mode rather than the softest one is essential, because the deepest mode need not be the one that condenses: in SrTiO3 the Γ ferroelectric mode is deeper than the R-point tilt yet is quantum-suppressed, and only the tilt drives the 105 K transition. (iii) Multi-mode SSCHA with an MLIP force engine is a clean gold standard for the martensitic bcc metals, but in its default deployment, with the free-energy Hessian truncated at the bubble, it systematically falsestabilises deep displacive (ferroelectric) instabilities and can diverge numerically. Theory predicts this truncation to fail for cubic metal halide perovskites; we measure the failure at benchmark scale and show it extends to oxide perovskites and, free of stochastic blowups, to cubic fluorites. The cheap soft-mode screen is therefore the more reliable finitetemperature indicator for the displacive instabilities that dominate generative crystalstructure-prediction (CSP) screening.
Related
- Dyna-Mat: End-to-end benchmarking of foundation machine learning interatomic potentials in finite-temperature ensembles
- Foundation interatomic potentials learn the practical chemistry they were never taught
- Leveraging neural network interatomic potentials for a foundation model of chemistry
- Analysis of fine-tuned foundation interatomic potentials and their knowledge distillation into a cluster expansion for Ni–Pt ordering thermodynamics
- Accelerating Mechanistic Understanding and Prediction of Chemical Selectivity with Foundation Machine-Learned Interatomic Potentials
BibTeX
@misc{cai2026finitetemperature,
title = {{Finite-temperature dynamic stability separates foundation machine-learning interatomic potentials that harmonic benchmarks rank equally}},
author = {Frank Cai},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008248/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008248/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Finite-temperature dynamic stability separates foundation machine-learning interatomic potentials that harmonic benchmarks rank equally
AU - Frank Cai
PY - 2026
DA - 2026/09/03
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008248/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008248/v1
AB - Foundation (universal) machine-learning interatomic potentials (MLIPs) are now used routinely as cheap stand-ins for DFT in high-throughput stability screening, but they are benchmarked almost exclusively against harmonic (0 K) phonons. Dynamic stability is physically a finite-temperature property, and an important class of materials, including cubic perovskites, bcc refractory metals and cubic fluorites, is harmonically unstable but thermally stabilised by anharmonicity. We test five foundation MLIPs (MACE-MP-0, CHGNet, ORB-v2, SevenNet-0, MatterSim) on this regime using a curated set of references whose finite-temperature behaviour is documented in the experimental and DFT literature, so that no new DFT ground truth is required. We introduce a cheap quantum selfconsistent-harmonic (SCHA) “soft-mode free energy” screen that evaluates every imaginary commensurate mode and calls the high-symmetry phase unstable if any of them condenses, and we cross-validate it against the gold-standard multi-mode stochastic SCHA (SSCHA). We report three findings together with a practical ensemble-disagreement guardrail. (i) At the harmonic level the models divide: MatterSim and SevenNet-0 reproduce every documented soft mode (accuracy 1.00), whereas MACE-MP-0 and CHGNet soften the bcc Zr/Hf instabilities to zero (15% false-stable rate) and ORB-v2 misses the SrTiO3 soft mode (8%). (ii) Harmonic accuracy does not predict finitetemperature accuracy in either direction: CHGNet is the worst model harmonically yet the second best at finite temperature, while MatterSim is harmonically perfect and only midtable, and on the matched set the two are non-significantly concordant at 100 K and significantly anti-associated at 300 K. Screening every imaginary mode rather than the softest one is essential, because the deepest mode need not be the one that condenses: in SrTiO3 the Γ ferroelectric mode is deeper than the R-point tilt yet is quantum-suppressed, and only the tilt drives the 105 K transition. (iii) Multi-mode SSCHA with an MLIP force engine is a clean gold standard for the martensitic bcc metals, but in its default deployment, with the free-energy Hessian truncated at the bubble, it systematically falsestabilises deep displacive (ferroelectric) instabilities and can diverge numerically. Theory predicts this truncation to fail for cubic metal halide perovskites; we measure the failure at benchmark scale and show it extends to oxide perovskites and, free of stochastic blowups, to cubic fluorites. The cheap soft-mode screen is therefore the more reliable finitetemperature indicator for the displacive instabilities that dominate generative crystalstructure-prediction (CSP) screening.
ER -