Machine learning interatomic potentials for ionic liquids and battery solvents

We develop machine learning force fields (MLFFs), based on the equivariant graph neural networks with NequIP/Allegro [1,2], for representative ionic liquids and conventional battery solvents. As the intermolecular interactions are subtle, and the dynamics of these electrolytes/solvents are quite slow, training a potential for these systems is not always straightforward. We develop a protocol for training MLFFs for complex, multicomponent solvents and electrolytes, which efficiently samples representative structures, to collect diverse, uncorrelated molecular configurations for training. This approach is shown to yield reliable simulations in the NVT ensemble, but not always in the NPT ensemble, where we find densities significantly lower than expected from our DFT calculations, similar to previous work. We develop an approach to remedy this issue, and test it on a number of electrolytes/solvents to ensure it is a robust method. In addition, we study the question of model transferability, the effect of long-range interactions and uncertainty of the model.

[1] S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nat. Commun., 13, 2453 (2022)

[2] A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. Owen, M. Kornbluth, and B. Kozinsky, Learning local equivariant representations for large-scale atomistic dynamics Nat. Commun., 14, 579 (2023)