Nanosecond MD of battery cathode materials with electron density description

Potentials derived with machine learning algorithms achieve the accuracy of high-fidelity quantum mechanical computations such as density functional theory (DFT), while allowing orders of magnitude lower computational time. In this work, we demonstrate the use of uncertainty aware equivariant graph neural networks for predicting spin-resolved electron densities, forces, and energies of the Na₃V₂(PO₄)₃ NASICON structured cathode. Due to the speedup in computational time, we are able to investigate structures of ∼300 atoms for 200 million timesteps. The ability to model larger systems on the nanosecond length scale with maintaining DFT level accuracy allowed critical insights into the diffusion characteristics of Na-ions, associated electron transfer processes, and dependence of diffusivity on sodium concentration in the structure.