Tensor-Field Molecular Dynamics - A Highly Accurate and Data-Efficient Interatomic Potential from SE(3)-equivariant Graph Neural Networks

We present Tensor-Field Molecular Dynamics (TFMD), a novel Deep Learning Interatomic Potential for accelerating Molecular Dynamics simulations. Our model uses SE(3)-equivariant convolutions over geometric tensors instead of the commonly used invariant convolutions over scalar features. We find that TFMD exhibits not only leading accuracy in the predicted atomic forces, but it also able to learn efficiently, outperforming even kernel-based methods on small data sets and opening the door to scalable simulations at beyond-DFT accuracy. We demonstrate our model on a diverse variety of systems, including organic molecules at DFT and CCSD(T) accuracy, water in different phases, a catalytic surface reaction, amorphous solids, and a superionic conductor. We show results from a series of dynamics simulations and demonstrate that TFMD can with high fidelity reproduce results from first-principles simulation and experiment.