Action-Minimization Meets Generative Modeling - Efficient Transition Path Sampling

Explore a groundbreaking approach to transition path sampling in molecular systems through this conference talk that bridges action minimization principles with generative modeling techniques. Learn how to leverage pre-trained generative models, specifically denoising diffusion and flow matching, to find probable paths connecting two points on energy landscapes without requiring expensive, task-specific training procedures. Discover how candidate paths can be interpreted as trajectories sampled from stochastic dynamics induced by learned score functions, transforming the challenge of finding high-likelihood transition paths into an optimization problem that minimizes the Onsager-Machlup action functional. Examine the zero-shot application of this methodology across varied molecular systems, demonstrating how it generates diverse, physically realistic transition pathways while generalizing beyond the original training datasets of pre-trained models. Understand the practical implications of this approach for drug discovery applications and how it can be seamlessly integrated into emerging generative models as they continue to scale with increased data availability, offering a more efficient alternative to traditional bespoke TPS models.