Analysis of fine-tuned foundation interatomic potentials and their knowledge distillation into a cluster expansion for Ni–Pt ordering thermodynamics
- 1Department of Mechanical Engineering, Yonsei University, 50 Yonsei-ro, Seodaemun-gu, Seoul, Republic of Korea, 03722
- 2Department of Mechanical Science and Engineering
- 3University of Illinois Urbana-Champaign
Abstract
Predicting alloy ordering thermodynamics requires the energies of thousands of atomic arrangements at fixed composition, far more than density functional theory (DFT) can routinely provide. Fine-tuned foundation machine-learning interatomic potentials (MLIPs) are a promising alternative, but whether they resolve the small energy differences between arrangements has not been established. Here, we examine this for fcc Ni–Pt in three steps. First, six foundation MLIPs fine-tuned on the same DFT data are evaluated on a 64-structure DFT benchmark. All six reach low global errors (2-20 meV/atom), but these errors are dominated by a composition-dependent offset that cancels between arrangements of one composition. We therefore evaluate the configurational energy Econf, the energy of a structure relative to the other structures of the same composition. After relaxation only four models reproduce DFT Econf (R2conf = 0.93-0.94), whereas fine-tuned eqV2 and CHGNet do not (0.48 and 0.37). Second, we locate the source of this failure. Linear probes on the frozen embeddings of all six models recover EconfDFT (R2 = 0.87 to 0.98), showing that the learned representations are intact. The accuracy is lost during relaxation: through the cell response of eqV2 and through inaccurate forces near the energy minimum in CHGNet. Third, we use the most accurate model, fine-tuned eqV3, as a teacher for knowledge distillation into a cluster expansion. Trained on 6,434 MLIP energies instead of 64 DFT energies, the distilled cluster expansion reduces the held-out DFT error from 4.43 to 2.79 meV/atom and reproduces the DFT-based transition temperatures within 167 K, comparable to the 120 K statistical uncertainty of the DFT-trained expansion itself. These results establish when fine-tuned foundation MLIPs can be trusted for alloy ordering and provide a route for transferring their accuracy to thermodynamic simulations.
Related
- A Lightweight Universal Machine-Learning Interatomic Potential via Knowledge Distillation for Scalable Atomistic Simulations
- Foundation interatomic potentials learn the practical chemistry they were never taught
- Leveraging neural network interatomic potentials for a foundation model of chemistry
- Dyna-Mat: End-to-end benchmarking of foundation machine learning interatomic potentials in finite-temperature ensembles
- Accelerating Mechanistic Understanding and Prediction of Chemical Selectivity with Foundation Machine-Learned Interatomic Potentials
BibTeX
@misc{hwang2026analysis,
title = {{Analysis of fine-tuned foundation interatomic potentials and their knowledge distillation into a cluster expansion for Ni–Pt ordering thermodynamics}},
author = {Jungjae Hwang and Jaejun Lee and Channyung Lee and Elif Ertekin and Kyoungmin Min},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009736/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009736/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Analysis of fine-tuned foundation interatomic potentials and their knowledge distillation into a cluster expansion for Ni–Pt ordering thermodynamics
AU - Jungjae Hwang
AU - Jaejun Lee
AU - Channyung Lee
AU - Elif Ertekin
AU - Kyoungmin Min
PY - 2026
DA - 2026/10/01
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009736/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009736/v1
AB - Predicting alloy ordering thermodynamics requires the energies of thousands of atomic arrangements at fixed composition, far more than density functional theory (DFT) can routinely provide. Fine-tuned foundation machine-learning interatomic potentials (MLIPs) are a promising alternative, but whether they resolve the small energy differences between arrangements has not been established. Here, we examine this for fcc Ni–Pt in three steps. First, six foundation MLIPs fine-tuned on the same DFT data are evaluated on a 64-structure DFT benchmark. All six reach low global errors (2-20 meV/atom), but these errors are dominated by a composition-dependent offset that cancels between arrangements of one composition. We therefore evaluate the configurational energy Econf, the energy of a structure relative to the other structures of the same composition. After relaxation only four models reproduce DFT Econf (R2conf = 0.93-0.94), whereas fine-tuned eqV2 and CHGNet do not (0.48 and 0.37). Second, we locate the source of this failure. Linear probes on the frozen embeddings of all six models recover EconfDFT (R2 = 0.87 to 0.98), showing that the learned representations are intact. The accuracy is lost during relaxation: through the cell response of eqV2 and through inaccurate forces near the energy minimum in CHGNet. Third, we use the most accurate model, fine-tuned eqV3, as a teacher for knowledge distillation into a cluster expansion. Trained on 6,434 MLIP energies instead of 64 DFT energies, the distilled cluster expansion reduces the held-out DFT error from 4.43 to 2.79 meV/atom and reproduces the DFT-based transition temperatures within 167 K, comparable to the 120 K statistical uncertainty of the DFT-trained expansion itself. These results establish when fine-tuned foundation MLIPs can be trusted for alloy ordering and provide a route for transferring their accuracy to thermodynamic simulations.
ER -