[Paper Review] A predictive machine learning force field framework for liquid electrolyte development

Sheng Gong, Yumin Zhang+14 more|arXiv (Cornell University)|Apr 10, 2024

Machine Learning in Materials Science8 citations

TL;DR

This paper introduces BAMBOO, a physics-informed, graph equivariant transformer–based ML force field framework for liquid electrolytes, with ensemble distillation and density alignment to predict density, viscosity, and ionic conductivity across diverse solvents and salts. It demonstrates transferability to unseen liquids and achieves state-of-the-art accuracy vs experimental data.

ABSTRACT

Despite the widespread applications of machine learning force fields (MLFF) in solids and small molecules, there is a notable gap in applying MLFF to simulate liquid electrolyte, a critical component of the current commercial lithium-ion battery. In this work, we introduce BAMBOO ( extbf{B}yteDance extbf{A}I extbf{M}olecular Simulation extbf{Boo}ster), a predictive framework for molecular dynamics (MD) simulations, with a demonstration of its capability in the context of liquid electrolyte for lithium batteries. We design a physics-inspired graph equivariant transformer architecture as the backbone of BAMBOO to learn from quantum mechanical simulations. Additionally, we introduce an ensemble knowledge distillation approach and apply it to MLFFs to reduce the fluctuation of observations from MD simulations. Finally, we propose a density alignment algorithm to align BAMBOO with experimental measurements. 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. The current model, trained on more than 15 chemical species, achieves the average density error of 0.01 g/cm$^3$ on various compositions compared with experiment.

Motivation & Objective

• Develop a transferable machine learning force field for complex liquid electrolytes that can learn from quantum data and generalize to unseen systems.
• Integrate physics-based components (electrostatics and dispersion) with a graph neural network backbone to accurately model liquids.
• Improve MD stability for MLFFs via ensemble knowledge distillation.
• Align MLFF predictions with experimental data through a density-alignment approach to enhance transferability.

Proposed method

• Propose a GET (graph equivariant transformer) based MLFF that partitions energy into semi-local, electrostatic, and dispersion contributions.
• Predict atomic energies, charges, and forces from atom types and relative coordinates; enforce Newton’s third law via pairwise force definitions.
• Compute electrostatic energy using predicted partial charges under a charge-equilibrium framework; incorporate D3(CSO) dispersion corrections.
• Train multiple GNNs with different random seeds on DFT-calculated energies/forces/charges and fuse them via ensemble knowledge distillation to stabilize MD.
• Align MLFF with experimental density data using a physics-based density alignment that links pressure adjustments to macroscopic observables and force refinements.
• Demonstrate transferability to unseen molecules and evaluate density, viscosity, and ionic conductivity across various solvents and salts.

Experimental results

Research questions

• RQ1 Can a single MLFF be trained to accurately describe diverse liquid electrolytes (solvents and salts) and transfer to unseen systems?
• RQ2 Do ensemble knowledge distillation and density alignment improve MD stability and agreement with experiments for liquid electrolytes?
• RQ3 How well does the BAMBOO framework predict density, viscosity, and ionic conductivity across a range of solvents and salt concentrations?
• RQ4 What is the impact of explicit electrostatic and dispersion treatments within the GET-based MLFF on prediction accuracy?
• RQ5 Can the framework discern local charge environments and solvation structures in Li-based electrolytes?

Key findings

• BAMBOO achieves average density error around 0.01 g/cm3 across diverse compositions.
• Viscosity and ionic conductivity predictions show strong agreement with experimental trends, with reported deviations of 17% and 26%, respectively, over multiple liquids.
• ensemble knowledge distillation reduces density prediction standard deviation by more than 50% across five models.
• Density alignment using a small experimental dataset reduces density error from ~0.05 to ~0.01 g/cm3 and improves related properties (viscosity, conductivity).
• The model demonstrates transferability to liquids not included in the DFT training set and can differentiate solvation motifs and partial charges based on local environments.