Micron-scale heterogeneous catalysis with Bayesian force fields from first principles and active learning

Abstract

Quantum-mechanically accurate reactive molecular dynamics (MD) at the scale of billions of atoms has been achieved for the heterogeneous catalytic system of H 2 /Pt(111) using the FLARE Bayesian force field. This achievement provides accelerated time-to-solution from first principles, with Bayesian active learning enabling efficient and autonomous training of the machine learning model. The resulting model is then deployed in LAMMPS on GPUs using the Kokkos performance portability library. The Bayesian force field provides quantitative uncertainty of predictions on every atomic environment, critical for detecting configurations in large reactive simulations that are outside of the training set. Scaling benchmarks were performed using realapplication MD of the H 2 /Pt(111) heterogeneous catalysis on the Summit supercomputer, with simulations reaching 0.5 trillion atoms on 4556 GPU nodes.

Towards predictive design of electrolyte solutions by accelerating ab initio simulation with neural networks

“…Therefore, a signicant amount of research effort is currently focused on developing improved, more sophisticated architectures to describe the potential energy surface, which in principle should be able to reduce the amount of required training data. [64][65][66][67][68][69][70][71][72][73][74] Many of these approaches incorporate new and exciting architectures known as graph neural networks (GNNs) or message passing neural networks (MPNNs). These approaches normally represent each atom as a multidimensional feature vector, which is a function of the atomic number and is iteratively updated using information about the distances and feature vectors of surrounding atoms.…”

“…Additionally, the linear scaling characteristic of these methods allows very large scale systems to be studied. 74,83 Although, additional effort will be needed to incorporate long range electrostatic and Casimir-Lifshitz forces into these large scale simulations. These methods hold particular promise for electrolyte solutions with slower dynamics where the long equilibration time makes direct AIMD particularly challenging, such as the organic electrolytes used in batteries, as well as water in salt electrolytes and ionic liquids.…”

Towards predictive design of electrolyte solutions by accelerating ab initio simulation with neural networks.

“…Therefore, a significant amount of research effort is currently focused on developing improved, more sophisticated architectures to describe the potential energy surface, which in principle should be able to reduce the amount of required training data. [64][65][66][67][68][69][70][71][72][73][74] Many of these approaches incorporate new and ex-citing architectures known as graph neural networks (GNNs) or message passing neural networks (MPNNs). These approaches normally represent each atom as a multidimensional feature vector, which is a function of the atomic number and is iteratively updated using information about the distances and feature vectors of surrounding atoms.…”

“…Additionally, the linear scaling characteristic of these methods allows very large scale systems to be studied. 74,83 Although, additional effort will be needed to incorporate long range electrostatic and Casimir-Lifshitz forces into these large scale simulations. These Fig.…”

Uncertainty Driven Active Learning of Coarse Grained Free Energy Models

“…However, in many cases the full GP can be mapped onto an exact model for predicting both the mean and uncertainty with a cost that is independent of the training set size. Application of these recent methods have been demonstrated to simulations of complex heterogeneous systems at record speed and size, reaching 500 billion atoms [51,60,61].…”