Machine Learning Force Field Framework for Liquid Electrolyte

Despite the widespread applications of machine learning force fields (MLFFs) in the modeling of solids and small molecules, a notable gap exists in their application to simulate liquid electrolytes, a critical component of commercial lithium-ion batteries. In this talk, we will first discuss the unique challenges associated with liquid electrolytes for atomistic modeling. We will then introduce BAMBOO (ByteDance AI Molecular Simulation Booster), a predictive framework designed for molecular dynamics (MD) simulations, showcasing its capabilities specifically within the context of liquid electrolytes for lithium batteries.
The core of BAMBOO is built upon a physics-inspired graph equivariant transformer architecture, which learns from quantum mechanical simulations. Additionally, we have pioneered an ensemble knowledge distillation approach, applying it to MLFFs to reduce fluctuations in MD simulation observations. To further align BAMBOO's predictions with experimental measurements, we propose a novel density alignment algorithm.
BAMBOO demonstrates state-of-the-art accuracy in predicting key electrolyte properties such as density, viscosity, and ionic conductivity across various solvents and salt combinations. Our current model, trained on data from over 15 chemical species, achieves an average density error of just 0.01 g/cm³ when compared to experimental data.
Furthermore, we will discuss the current limitations of MLFFs in simulating organic electrolytes and the opportunities for further advancements, particularly in enhancing inference speed and improving transferability to a broader range of molecules.