Reinforcement learning amortizes transition-state physics into one-step flow models
Abstract
Transition-state searches remain a major bottleneck in reaction discovery, as identifying valid saddle-point structures requires numerous expensive quantum-chemical calculations. Generative models can reduce this burden by proposing candidates from reactant and product geometries, but supervised training on geometries alone does not enforce the physical conditions required of a transition state, while enforcing them during sampling is costly. We introduce Trèfle, a one-step flow model that amortizes physics-guided generation into training: a mean-flow objective replaces multi-step sampling, a low-cost surrogate replaces density functional theory (DFT) reward evaluations, and reinforcement-learning post-training with rewards that enforce those conditions replaces inference-time guidance. Trèfle generates accurate transition states at 30× the throughput of multi-step baselines. It transfers few-shot to ten unseen transition metals and generalizes zero-shot to molecules far larger than any in training. With DFT refinement and reaction-path validation, Trèfle-seeded searches recover all 13 unseen γ-ketohydroperoxide channels, while post-training raises per-attempt recovery from 28% to 50% across 85 unseen bimolecular channels. Trèfle thus provides fast, reliable seeds for automated reaction discovery.
Related
- Development of a Deep Reinforcement Learning Approach for Transition State Search in Molecular Systems: A Transferable Model Across Structural and Dimensional Variations
- Fine-tuning universal machine learning potentials for transition state search in surface catalysis
- Learning continuous reaction paths for transition-state prediction
- Physics-Informed Explainable Machine Learning for Hybrid State of Health and Remaining Useful Life Estimation of Lithium-Ion Batteries
- Physics-Aligned Electronic Ground-State Learning Improves Generalization
BibTeX
@misc{li2026reinforcement,
title = {{Reinforcement learning amortizes transition-state physics into one-step flow models}},
author = {Yunyang Li and Zechang Sun and Kuang Yu and Wen Yan and Mark Gerstein and Hung Q. Pham},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.32073},
eprint = {2609.32073},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.32073},
}
RIS
TY - UNPB
TI - Reinforcement learning amortizes transition-state physics into one-step flow models
AU - Yunyang Li
AU - Zechang Sun
AU - Kuang Yu
AU - Wen Yan
AU - Mark Gerstein
AU - Hung Q. Pham
PY - 2026
DA - 2026/09/25
PB - arXiv
DO - 10.48550/arxiv.2609.32073
UR - https://arxiv.org/abs/2609.32073
AB - Transition-state searches remain a major bottleneck in reaction discovery, as identifying valid saddle-point structures requires numerous expensive quantum-chemical calculations. Generative models can reduce this burden by proposing candidates from reactant and product geometries, but supervised training on geometries alone does not enforce the physical conditions required of a transition state, while enforcing them during sampling is costly. We introduce Trèfle, a one-step flow model that amortizes physics-guided generation into training: a mean-flow objective replaces multi-step sampling, a low-cost surrogate replaces density functional theory (DFT) reward evaluations, and reinforcement-learning post-training with rewards that enforce those conditions replaces inference-time guidance. Trèfle generates accurate transition states at 30× the throughput of multi-step baselines. It transfers few-shot to ten unseen transition metals and generalizes zero-shot to molecules far larger than any in training. With DFT refinement and reaction-path validation, Trèfle-seeded searches recover all 13 unseen γ-ketohydroperoxide channels, while post-training raises per-attempt recovery from 28% to 50% across 85 unseen bimolecular channels. Trèfle thus provides fast, reliable seeds for automated reaction discovery.
ER -