By accelerating molecular dynamics with machine-learning, we can predict the real-time motion of thousands of atoms for microsecond times, at first-principles accuracy. This capability allows us to study the evolution of  bimetallic surfaces and nanoparticles used as catalysts for chemical production. These simulations reveal the mechanisms of surface restructuring and explain experimental observations.

https://doi.org/10.1021/jacs.0c06401

https://www.seas.harvard.edu/news/2020/10/close-and-personal-catalysts 

https://www.youtube.com/watch?v=O53b_Cfz06Y 

Explicit heterogeneous reactive dynamics simulations at large scale are important to discover reaction pathways and explore the effects of surface composition and structure. But these investigations have been challenging due to the lack of accurate fast models. By leveraging uncertainty-driven active learning in FLARE, we train models for describing surface reactions.  In a prototypical hydrogen activation reaction on Pt surface, we are able to train models automatically in about a week of wall time and deploy the model on supercomputers, describing billions of atoms at near-quantum accuracy and in excellent agreement with experimental reactivity measurements.

https://arxiv.org/abs/2106.01949

https://arxiv.org/abs/2204.12573