Data-Efficient Machine Learning Potentials for Organic Reactions through Active-Learning-Guided Fine-Tuning
Abstract
Transition states (TSs) are central to reaction barriers and mechanisms, yet their accurate localization remains computationally demanding. Reactive machine learning potentials can accelerate TS optimization by rapidly predicting potential energy surfaces, but achieving chemical transferability requires extensive sampling and quantum chemical (QC) labeling of reactive conformations. Here, we introduce an active-learning-guided fine-tuning strategy that couples pretrained chemical knowledge for broad generalization with adaptive labeling of informative reactive conformations for data efficiency. The resulting MACEPre-DeePEST-OS enables efficient TS optimization and barrier predictions across diverse organic reactions while requiring QC labeling for only ~2% of approximately 19 million candidate conformations. Across three in-distribution benchmarks, MACEPre-DeePEST-OS achieves an average root mean square deviation of 0.080 Å for TS geometries and a mean absolute error (MAE) of 0.938 kcal/mol for barriers. In cross-dataset evaluation on an external OrgReact subset, fine-tuning reduces the barrier MAE by 58% relative to the pretrained model. Out-of-distribution testing on a Ugi reaction network containing structures with more than 70 atoms yields a barrier MAE of 1.09 kcal/mol, indicating transfer to molecular sizes and reaction complexity beyond the fine-tuning data. These results establish a data-efficient route toward transferable reactive potentials for quantitative analysis of reaction mechanisms and competing pathways.
Related
- Efficient and Transferable Machine Learning Potentials for Zn-Coordinated Zeolitic Imidazolate Frameworks Through Structural Diversity Sampling and Transfer Learning.
- FACTO: fast atom-centered tensor operator for accurate and efficient foundational 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
BibTeX
@misc{ren2026dataefficient,
title = {{Data-Efficient Machine Learning Potentials for Organic Reactions through Active-Learning-Guided Fine-Tuning}},
author = {Kaipai Ren and Yujing Zhao and Kun Tang and Juntao Wang and Guangran Zhang and Qilei Liu},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009333/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009333/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Data-Efficient Machine Learning Potentials for Organic Reactions through Active-Learning-Guided Fine-Tuning
AU - Kaipai Ren
AU - Yujing Zhao
AU - Kun Tang
AU - Juntao Wang
AU - Guangran Zhang
AU - Qilei Liu
PY - 2026
DA - 2026/09/23
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009333/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009333/v1
AB - Transition states (TSs) are central to reaction barriers and mechanisms, yet their accurate localization remains computationally demanding. Reactive machine learning potentials can accelerate TS optimization by rapidly predicting potential energy surfaces, but achieving chemical transferability requires extensive sampling and quantum chemical (QC) labeling of reactive conformations. Here, we introduce an active-learning-guided fine-tuning strategy that couples pretrained chemical knowledge for broad generalization with adaptive labeling of informative reactive conformations for data efficiency. The resulting MACEPre-DeePEST-OS enables efficient TS optimization and barrier predictions across diverse organic reactions while requiring QC labeling for only ~2% of approximately 19 million candidate conformations. Across three in-distribution benchmarks, MACEPre-DeePEST-OS achieves an average root mean square deviation of 0.080 Å for TS geometries and a mean absolute error (MAE) of 0.938 kcal/mol for barriers. In cross-dataset evaluation on an external OrgReact subset, fine-tuning reduces the barrier MAE by 58% relative to the pretrained model. Out-of-distribution testing on a Ugi reaction network containing structures with more than 70 atoms yields a barrier MAE of 1.09 kcal/mol, indicating transfer to molecular sizes and reaction complexity beyond the fine-tuning data. These results establish a data-efficient route toward transferable reactive potentials for quantitative analysis of reaction mechanisms and competing pathways.
ER -