Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics

Explore an innovative approach to accelerating molecular dynamics simulations through machine learning in this 1 hour 15 minute talk. Dive into the Timewarp method, which utilizes normalizing flows as proposal distributions in Markov chain Monte Carlo simulations targeting the Boltzmann distribution. Learn how this transferable algorithm can generalize to unseen small peptides, providing significant wall-clock acceleration compared to standard molecular dynamics. Discover the potential of this technique to simulate molecular dynamics over extended timescales, addressing challenges in studying important processes like binding and folding. Gain insights into the method's implementation, including conditional normalizing flows, atom transformers, kernel self-attention, and augmented MCMC. Understand the validation process for new metastable states and engage with the speakers during the Q&A session to deepen your understanding of this cutting-edge approach in computational chemistry and drug discovery.

- Intro
- Molecular dynamics
- Datasets
- Conditional noarmalizing flows
- Atom transformer and kernel self-attention
- Augmented MCMC
- Batching the timewarp MCMC algorithm
- Training trajectories
- Timewarp MCMC
- Validation of new metastable states
- Summary
- Q+A