BAMBOO: a predictive and transferable machine learning force field framework for liquid electrolyte development

Liquid electrolyte is an indispensable component in most of electrochemical energy devices that include, but not limited to lithium ion and lithium metal batteries [1, 2, 3]. The existing commercial electrolytes are primarily carbonate-based, and it is common to find a commercial electrolyte composed of more than five, even up to ten different components to meet various aspects of cell performances. Recent developments have expanded the electrolyte designs to high-concentrated [4], localized high-concentrated [5], and fluorinated ether-based electrolytes [6, 7]. These novel designs aim to engineer molecular-level solvation structures for improved solvation/desolvation [8] performance, solid electrolyte interphase [9], and electrochemical stability [10]. Experimentally exploring molecular interactions for rational design is costly, time-consuming, and heavily reliant on chemists' intuition and experience. These limitations pose challenges in transitioning from proof of concept in a lab to commercialization, particularly due to the exponential complexity involved in optimizing properties and local solvation structures for multi-component liquid electrolyte systems. Atomistic simulations offer an efficient and flexible alternative to exhaust experimentation. They can accurately capture the evolving ion-solvent polarizable interactions, thereby, providing reliable bulk and molecular level property predictions. However, requirements such as sufficient simulation time and scale need to be met.