Explore a 41-minute conference talk on differentiable simulations for enhanced sampling of rare events in chemical reactions. Dive into the innovative approach of using differentiable simulations (DiffSim) to discover and sample chemical transformations without relying on preselected collective variables. Learn how path-integral optimization merges reaction path discovery and biasing potential estimation into a single end-to-end problem. Discover improvements to standard DiffSim, including partial backpropagation and graph mini-batching, that enhance training stability and efficiency. Examine the successful application of DiffSim in discovering transition paths for the Muller-Brown model potential and alanine dipeptide. Gain insights into the challenges of molecular dynamics simulations of chemical reactions, biased Langevin dynamics, and training in 2D cases. Explore future outlooks and participate in a Q&A session to deepen your understanding of this cutting-edge approach in computational chemistry and drug discovery.
Overview
Syllabus
- Intro
- Differentiable Simulations
- The Challenge of MD Simulation of Chemical Reactions
- Biased Langevin Dynamics
- 2D Case: Training
- Concave Surfaces
- Future Outlooks
- Q+A
Taught by
Valence Labs