Abstract

Electrolytes, such as those based on ionic liquids (ILs) or carbonate solvents, are typically simulated using classical force fields because they are computational cheap enough to reach the timescales and system sizes required. These force fields are often not quantitatively accurate when comparing to experiments, however, which motivates the development of machine learning interatomic potentials (MLIPs) to obtain more reliable results. However, MLI Ps are often only trained and deployed on a single composition, but electrolytes are typically complex mixtures, and we want to study their physiochemical properties as a function of salt concentration, different solvents and additives, for example. We demonstrate a method to train MLI Ps to be chemically transferable, using a salt-in-ionic liquid as a test case. We find that to learn the energy to be transferable, a minimum number of compositions are required, and we generalise this result to provide guidelines for investigating more complex systems [1]. Simulating electrolytes is also typically performed in the NPT ensemble, but it has been found that training MLI Ps to be stable in the NPT ensemble can be difficult. We investigated this problem for various systems, including ILs and carbonates solvents, and found that Allegro could be made to be stable for these systems, with the chosen hyperparameters and datasets being key factors in determining their stability. [1] – Goodwin et al. arXiv:2403.01980 (2024)