Machine Learning of Force Fields for Li-Ion Conductors

In this work, we evaluate multiple state-of-the-art graph neural networks and compare their performance on the Li-ion molecular dynamics tasks with the goal of accurately prediction ionic diffusivity and conductivity values. Apart from accuracy of predictions of energies and forces, we add the task of predicting the RDF after a molecular dynamics run, as well as the conductivity per se.<br/>We start by comparing three materials with similar composition but different crystal structure and conductivities and chose the most transferable model. Surprisingly, we find that equivariance does not necessarily improve the model.<br/>Secondly, we find that the majority of pretrained models (e.g. OC22 or MACE) fail dramatically on the Li ion conduction. We thus create our own dataset of 1000 Li-containing materials and train our model on it.