Machine Learning Gradients in Molecular Simulations

Explore the intersection of machine learning and molecular simulations in this 57-minute talk by Rafael Gomez-Bombarelli from MIT. Delve into the crucial role of gradient-based optimization and differentiable programming in deep learning, particularly in scientific applications. Discover how merging machine learning models with physics-based simulators can enhance expensive simulations and bridge the gap between incomplete models and experimental data. Learn about research examples showcasing the exploitation of ML surrogate functions and their gradients in molecular simulations. Examine applications including active learning of machine learning potentials, adversarial attacks on differentiable uncertainty, data-driven collective variables for enhanced sampling, coarse-graining and backmapping all-atom simulations, and interpolation of differentiable alchemical atom types for thermodynamic integration.