Entropy-Guided Dataset Reduction for Machine-Learned Interatomic Potentials
- 1Department of Chemical Engineering, University of Michigan, 500 S State St, Ann Arbor, MI, USA, 48109
- 2Materials Science Division
- 3Lawrence Livermore National Laboratory
Abstract
Machine-learned interatomic potentials (MLIPs) deliver near-quantum-mechanical accuracy at a fraction of the computational cost. Their performance depends on the quality and composition of the training data: the training set must cover the range of atomic environments the model will encounter, yet unnecessary redundancy increases training cost without improving accuracy and can inadvertently introduce training biases. Dataset pruning (i.e., selecting a compact, representative subset from a larger collection) has emerged as a practical strategy for building high-quality training sets. Recently, methods based on informationtheoretic diversity measures such as QUESTS (Quick Uncertainty and Entropy from STructural Similarity) have shown significant promise, however, QUESTS requires specification of a lengthscale parameter (kernel bandwidth) that controls how similar two atomic environments must be before one is considered redundant. We show that operationally useful values of this parameter vary across structurally distinct phases, making naive automated pruning of heterogeneous datasets challenging. We present a pruning framework that addresses this problem by adapting the bandwidth independently for distinct thermodynamic subsets of the training dataset using an automatable elbow criterion applied to QUESTS informationentropy metrics. Farthest-point sampling (FPS) is then applied within each group to select maximally diverse atomic environments. Using the Small Carbon ChIMES 2.0 dataset comprising diamond, graphite, and liquid carbon spanning a broad range of thermodynamic conditions, we show that use of a single global bandwidth yields pruned datasets that can be strongly dominated by a single phase, particularly when the distribution of environments varies significantly between phases (e.g., liquid vs crystalline).1,2 Relative to global pruning at the same retained fraction, group-aware adaptive pruning reduces test force-MAE by up to 37% for diamond and 26% for graphite for models trained on only 5% of the original training data, with corresponding improvements in radial distribution function (RDF) fidelity. Benefits are smaller in structurally heterogeneous conditions where a single thermodynamic label encompasses multiple distinct local bonding environments with different characteristic length scales, indicating that finer grouping is the natural remedy in such regimes. Across crystalline conditions, the elbow-selected bandwidth aligns with the onset of rapid error growth as the dataset is pruned, providing a practical, model-agnostic guide for pruning decisions that requires no model training.
Related
- SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials
- Correcting the Liquid Densities of Universal Machine-Learning Interatomic Potentials Using Local-Number-Density Repulsion
- pdb2reaction: End-to-End Reaction-Path Elucidation from PDB Structures Using Machine-Learning Interatomic Potentials
- Applicability Domain of Universal Machine Learning Interatomic Potentials at Solid-State Battery Interfaces
- Accurate and Rapid Vibrational Spectra of Molecular Crystals from Machine-Learned Interatomic Potentials
BibTeX
@misc{laubach2026entropyguided,
title = {{Entropy-Guided Dataset Reduction for Machine-Learned Interatomic Potentials}},
author = {Benjamin R. Laubach and Joshua A. Vita and Kyle Bushick and Logan Williams and Vincenzo Lordi and Rebecca K. Lindsey},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008542/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008542/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Entropy-Guided Dataset Reduction for Machine-Learned Interatomic Potentials
AU - Benjamin R. Laubach
AU - Joshua A. Vita
AU - Kyle Bushick
AU - Logan Williams
AU - Vincenzo Lordi
AU - Rebecca K. Lindsey
PY - 2026
DA - 2026/09/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008542/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008542/v1
AB - Machine-learned interatomic potentials (MLIPs) deliver near-quantum-mechanical accuracy at a fraction of the computational cost. Their performance depends on the quality and composition of the training data: the training set must cover the range of atomic environments the model will encounter, yet unnecessary redundancy increases training cost without improving accuracy and can inadvertently introduce training biases. Dataset pruning (i.e., selecting a compact, representative subset from a larger collection) has emerged as a practical strategy for building high-quality training sets. Recently, methods based on informationtheoretic diversity measures such as QUESTS (Quick Uncertainty and Entropy from STructural Similarity) have shown significant promise, however, QUESTS requires specification of a lengthscale parameter (kernel bandwidth) that controls how similar two atomic environments must be before one is considered redundant. We show that operationally useful values of this parameter vary across structurally distinct phases, making naive automated pruning of heterogeneous datasets challenging. We present a pruning framework that addresses this problem by adapting the bandwidth independently for distinct thermodynamic subsets of the training dataset using an automatable elbow criterion applied to QUESTS informationentropy metrics. Farthest-point sampling (FPS) is then applied within each group to select maximally diverse atomic environments. Using the Small Carbon ChIMES 2.0 dataset comprising diamond, graphite, and liquid carbon spanning a broad range of thermodynamic conditions, we show that use of a single global bandwidth yields pruned datasets that can be strongly dominated by a single phase, particularly when the distribution of environments varies significantly between phases (e.g., liquid vs crystalline).1,2 Relative to global pruning at the same retained fraction, group-aware adaptive pruning reduces test force-MAE by up to 37% for diamond and 26% for graphite for models trained on only 5% of the original training data, with corresponding improvements in radial distribution function (RDF) fidelity. Benefits are smaller in structurally heterogeneous conditions where a single thermodynamic label encompasses multiple distinct local bonding environments with different characteristic length scales, indicating that finer grouping is the natural remedy in such regimes. Across crystalline conditions, the elbow-selected bandwidth aligns with the onset of rapid error growth as the dataset is pruned, providing a practical, model-agnostic guide for pruning decisions that requires no model training.
ER -