Deep Generative Models for Molecular Simulation

Go back to programme

Deep Generative Models for Molecular Simulation

  • Watch

    * Register (or log in) to the AI4G Neural Network to add this session to your agenda or watch the replay

    • Register

      Watch

      * Register (or log in) to the AI4G Neural Network to add this session to your agenda or watch the replay

    This session will explore advanced methods in deep generative modeling for molecular simulations, focusing on a novel approach to improve the accuracy and efficiency of normalizing flows in approximating complex target distributions, such as Boltzmann  distributions of physical systems. Traditional methods for training flows often struggle with mode-seeking behavior, use samples from the target generated beforehand by expensive MCMC methods, or use stochastic losses that have high variance. reliance on costly MCMC-generated samples, or high-variance stochastic losses.  

    We will introduce Flow AIS Bootstrap (FAB), a new technique that integrates annealed importance sampling (AIS) to overcome these limitations. By minimizing the mass-covering α-divergence with α=2, FAB reduces importance weight variance and generates samples in regions where the flow poorly approximates the target, aiding in the discovery of new modes.  

    The session will highlight how FAB accurately approximates multimodal targets, outperforming previous methods. To the best of our knowledge, we are the first to learn the Boltzmann distribution of the alanine dipeptide molecule using only the unnormalized target density, without access to samples generated via Molecular Dynamics (MD) simulations: FAB produces better results than training via maximum likelihood on MD samples while using 100 times fewer target evaluations. After reweighting the samples, we obtain unbiased histograms of dihedral angles that are almost identical to the ground truth.

    Share this session

    Are you sure you want to remove this speaker?