[Paper Review] Micron-scale heterogeneous catalysis with Bayesian force fields from first principles and active learning

Anders Johansson, Yu Xie+5 more|arXiv (Cornell University)|Apr 26, 2022

Machine Learning in Materials Science23 citations

TL;DR

The authors develop FLARE, a Bayesian many-body force field with active learning to perform first-principles-level reactive MD of H2 on Pt(111) at micron scales, achieving up to 0.5 trillion atoms on Summit with uncertainty quantification and autonomous training.

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 real-application MD of the H$_2$/Pt(111) heterogeneous catalysis on the Summit supercomputer, with simulations reaching 0.5 trillion atoms on 4556 GPU nodes.

Motivation & Objective

• Address the challenge of simulating heterogeneous catalysis at micrometer scales with quantum-mechanical accuracy.
• Develop a fast, scalable, uncertainty-aware ML force field trained from first principles.
• Enable autonomous data acquisition via Bayesian active learning to efficiently explore configuration space.
• Demonstrate large-scale application to H2/Pt(111) with realistic reaction conditions on leadership-class HPC.

Proposed method

• Use FLARE Bayesian force field with many-body ACE-B2 descriptors and Gaussian process regression.
• Employ a mapped sparse GP to predict energies/forces/stress and per-atom uncertainties with O(n_d^2) cost for mean and variance.
• Apply Bayesian active learning: MD with uncertainty monitoring triggers DFT calculations to enrich training data.
• Integrate with LAMMPS on GPUs via Kokkos for high-performance, scalable MD.
• Scale to micrometer-length, billion-atom systems and optimize for heterogeneous environments with descriptor batching and GPU acceleration.

Experimental results

Research questions

• RQ1 Can a Bayesian, uncertainty-aware ML force field reproduce first-principles accuracy for H/Pt heterogeneous catalysis across large scales?
• RQ2 What data-efficiency and active-learning strategies are required to train such a model with minimal DFT calls?
• RQ3 How does the model perform in terms of speed, scalability, and reliability on leadership-class HPC for micrometer-scale reactive MD?
• RQ4 Do simulations reproduce experimentally relevant reaction kinetics and activation barriers for H2 dissociation and recombination on Pt(111)?

Key findings

• FLARE achieves 10.5 million atom-steps per second per GPU node, outperforming state-of-the-art MLFFs on similar tasks.
• Production-scale simulations reach 0.5 trillion atoms on 4556 Summit GPU nodes, with near-perfect weak scaling to 40B atoms.
• Uncertainty quantification is integrated with negligible speed impact, enabling on-the-fly detection of unfamiliar configurations.
• Bayesian active learning reduced training data needs dramatically, requiring 575 DFT calls versus AIMD-scale data generation.
• Mean absolute errors against DFT for energies/forces/stress are 1.7 meV/atom, 91 meV/Å (Pt)/74 meV/Å (H), and 0.6 meV/Å^3, respectively, outperforming ReaxFF.
• Reaction-rate estimates (activation energy) closely match experiment (0.25 eV vs 0.23 eV).