Learning continuous reaction paths for transition-state prediction
Abstract
Transition states are defined by reaction pathways, yet most machine-learning methods predict them as isolated geometries. We introduce MARC-TS, a two-stage framework that learns a continuous, endpoint-conditioned path, queries it at any resolution and uses local path context to refine a transition-state candidate. We construct T1x-IRC-8K, a dataset of 8,209 reactions and 1,088,725 path-resolved geometries. On held-out reactions, the path model reduced complete-path error by 48.4% relative to endpoint interpolation, and the localizer achieved a mean aligned structural error of 0.127 Å. Quantum-chemical optimization and vibrational analysis yielded 405 frequency-confirmed first-order saddle-point candidates from 410 predictions. In a 100-reaction nudged elastic band comparison, learned-path initialization reached a joint geometry-and-force target for 66% of reactions, compared with 12% for geometric interpolation after 100 optimizer steps. By treating the path as a reusable representation rather than an auxiliary output, MARC-TS connects transition-state prediction, mechanistic interpretation and quantum-chemical refinement.
Related
- IMPLICIT LEARNING: REACTION OUTCOME PREDICTION FROM POSITIVE-ONLY DATA
- Bridging the Literature-to-Laboratory Gap with Language Models and Active Learning for Reaction Feasibility Prediction and Autonomous Exploration
- Machine Learning for High-Throughput Reaction Yield Prediction in DNA-Encoded Library Synthesis
- All Paths All at Once: A Quantum Monte Carlo and Neural Committor Approach to Reaction Pathway Search
- Transition states and dynamical bottlenecks in the Li-NC <-> Li-CN isomerizing reaction with the Lagrangian betweenness
BibTeX
@misc{yang2026learning,
title = {{Learning continuous reaction paths for transition-state prediction}},
author = {Yexiang Yang and Linlin Zhong},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.25523},
eprint = {2609.25523},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.25523},
}
RIS
TY - UNPB
TI - Learning continuous reaction paths for transition-state prediction
AU - Yexiang Yang
AU - Linlin Zhong
PY - 2026
DA - 2026/09/22
PB - arXiv
DO - 10.48550/arxiv.2609.25523
UR - https://arxiv.org/abs/2609.25523
AB - Transition states are defined by reaction pathways, yet most machine-learning methods predict them as isolated geometries. We introduce MARC-TS, a two-stage framework that learns a continuous, endpoint-conditioned path, queries it at any resolution and uses local path context to refine a transition-state candidate. We construct T1x-IRC-8K, a dataset of 8,209 reactions and 1,088,725 path-resolved geometries. On held-out reactions, the path model reduced complete-path error by 48.4% relative to endpoint interpolation, and the localizer achieved a mean aligned structural error of 0.127 Å. Quantum-chemical optimization and vibrational analysis yielded 405 frequency-confirmed first-order saddle-point candidates from 410 predictions. In a 100-reaction nudged elastic band comparison, learned-path initialization reached a joint geometry-and-force target for 66% of reactions, compared with 12% for geometric interpolation after 100 optimizer steps. By treating the path as a reusable representation rather than an auxiliary output, MARC-TS connects transition-state prediction, mechanistic interpretation and quantum-chemical refinement.
ER -