Predicting transition states (TS) in chemical reactions is essential for uncovering reaction mechanisms. Most existing TS predictors rely on linear flow matching, assuming a straight reaction path that rarely matches the true molecular trajectory.
ReCurveflow proposes a flow‑matching framework that supervises on continuously curved reference paths interpolated from a full NEB (Nudged Elastic Band) band of geometries. The model learns a velocity field that drives a given reactant geometry along the curved path toward the TS geometry.
To improve inference robustness, ReCurveflow introduces an off‑path correction mechanism. When the generated geometry deviates from the reference trajectory, the model emits a corrective velocity field, mitigating exposure bias and enhancing TS prediction accuracy.
Across three data splits and six evaluation metrics, ReCurveflow outperforms seven baselines on the majority of split‑metric combinations. Qualitative analysis shows that the generated energy profiles closely follow the NEB reference, the provided initial structures alleviate the NEB optimization bottleneck, and the learned velocity fields exhibit the intended corrective behavior.
The codebase is publicly available at https://github.com/dmis-lab/ReCurveflow.
Blogger's Review: By leveraging curved reference trajectories and off‑path correction, ReCurveflow overcomes the limitations of linear flow matching, offering a more realistic approach to TS prediction. It represents a promising direction at the intersection of computational chemistry and machine learning.