(8b) Methods and Applications of Ex-Machina Molecular Dynamics to Catalytic Processes

We pursue the paradigm of “ex-machina” computations where data-driven approximations are automatically developed using machine learning algorithms and enable access to dynamics of complex chemical systems. Non-parametric regression methods allow for learning of potential energy surfaces from expensive quantum calculations. To accelerate molecular dynamics calculations, we developed the Neural equivariant interatomic potential model (NequIP) based on tensor-valued symmetry-preserving layer architectures and used them to achieve state-of-the-art accuracy and training efficiency for simulating dynamics of molecules, water and heterogeneous catalysts. In order to enable autonomous selection of the training set for reactive systems, we developed the FLARE adaptive closed-loop algorithm that constructs accurate and uncertainty-aware Bayesian force fields on-the-fly from a molecular dynamics simulation, using Gaussian process regression. We demonstrate the performance of ML-accelerated MD simulations by studying bimetallic catalyst surface restructuring and HD exchange reactions. Finally, we develop dimensionality reduction techniques in order to automatically identify the reaction coordinates from dynamics simulations, that can be used to enhance sampling of rare transitions and to estimate reaction rates.