Machine Learning Force-Fields promise to move past the dilemma of obtaining interatomic force-fields that are both highly accurate while also computationally efficient. Graph Neural Networks in particular have in recent years emerged as a promising approach for learning representation from raw atomic data without the need for hand-crafted desciptors. While most Graph Neural Networks leverage invariant convolutions over scalars, we present a Graph Neural Network that leverages SE(3)-equivariant convolutions over tensors instead. We demontrate that the proposed network achieves state-of-the-art accuracy on a wide variety of systems, including small, organic molecules, water in different phases, a chemical reaction, an amorphous solid, and a Li superionic conductor. Finally, we show that NequIP can predict structural and kinetic properties from Molecular Dynamics simulations with high fidelity in comparison to ab-initio simulation.

NequIP - leveraging SE(3)-equivariant graph convolutions for accurate and sample-efficient machine learning force-fields

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.