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 Tr\`efle, 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. Tr\`efle generates accurate transition states at $30\times$ 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, Tr\`efle-seeded searches recover all 13 unseen $\gamma$-ketohydroperoxide channels, while post-training raises per-attempt recovery from 28% to 50% across 85 unseen bimolecular channels. Tr\`efle thus provides fast, reliable seeds for automated reaction discovery.