Full-data accuracy with fewer labels for training and fine-tuning machine-learning force fields
Abstract
Machine-learning force fields (MLFFs) are reliable only near their training distribution, making efficient construction of diverse training sets a major bottleneck for both train-from-scratch and foundation fine-tuning workflows. Active learning can reduce this cost, but standard model-committee uncertainty is impractical for foundation MLFFs because each committee member requires a separate fine-tuning run. We present an active-learning workflow based on last-layer-projection regression (LLPR), a forward-pass-cheap per-configuration uncertainty estimator. Across molecular, condensed-phase, and electrolyte systems, LLPR identifies compact, high-value training sets that recover full-data accuracy using only a small fraction of electronic-structure labels. In foundation-model fine-tuning, LLPR-selected configurations approach the practical fine-tuning ceiling with substantially fewer labels than random selection. In iterative electrolyte fine-tuning, LLPR detects unphysical local coordination before new labelling, calibrates uncertainty in force-error units, and supports retrospective convergence monitoring. The resulting models reproduce reference density and Li--O coordination structure, providing a scalable uncertainty-quantification strategy across MLFF training regimes.
Related
- Accuracy, Reliability, and Efficiency of a MACE Machine-Learning Interatomic Potential for Aspirin: A Reproducible Benchmark for Molecular Simulation
- Fine-tuning universal machine learning potentials for transition state search in surface catalysis
- Data-Efficient Machine Learning Potentials for Organic Reactions through Active-Learning-Guided Fine-Tuning
- Targeted fine-tuning of machine-learning interatomic potentials for phonons and phase transitions.
- Training of Neural Network Machine Learning Potentials for Activation Barrier Calculations in Predicting Growth Modes of Deposited Metals
BibTeX
@misc{bi2026fulldata,
title = {{Full-data accuracy with fewer labels for training and fine-tuning machine-learning force fields}},
author = {Sheng Bi and Yi-Ze Wang and Jun Cheng},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2607.14486},
eprint = {2607.14486},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.14486},
}
RIS
TY - UNPB
TI - Full-data accuracy with fewer labels for training and fine-tuning machine-learning force fields
AU - Sheng Bi
AU - Yi-Ze Wang
AU - Jun Cheng
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2607.14486
UR - https://arxiv.org/abs/2607.14486
AB - Machine-learning force fields (MLFFs) are reliable only near their training distribution, making efficient construction of diverse training sets a major bottleneck for both train-from-scratch and foundation fine-tuning workflows. Active learning can reduce this cost, but standard model-committee uncertainty is impractical for foundation MLFFs because each committee member requires a separate fine-tuning run. We present an active-learning workflow based on last-layer-projection regression (LLPR), a forward-pass-cheap per-configuration uncertainty estimator. Across molecular, condensed-phase, and electrolyte systems, LLPR identifies compact, high-value training sets that recover full-data accuracy using only a small fraction of electronic-structure labels. In foundation-model fine-tuning, LLPR-selected configurations approach the practical fine-tuning ceiling with substantially fewer labels than random selection. In iterative electrolyte fine-tuning, LLPR detects unphysical local coordination before new labelling, calibrates uncertainty in force-error units, and supports retrospective convergence monitoring. The resulting models reproduce reference density and Li--O coordination structure, providing a scalable uncertainty-quantification strategy across MLFF training regimes.
ER -