Predictive Supremacy of Chemical Foundational Model for Battery Electrolytes

Abstract

Electrolytes are a critical design black-box for batteries that aspects all major performance metrices. Electrolyte discovery and optimization can accelerate decarbonization and has drawn major focus from across all economic sectors. Consequently, electrolyte formulations are now guarded trade secrets of battery companies that are rarely shared publicly, making electrolyte data landscape scarce for AI-driven discovery. The most challenging aspect of development of battery electrolytes is their non-generalizability. There exist several major battery chemistries in research stages that require unique concoction of solvent-salt systems with special attention to underlying phenomenons such as lithium solvation, surface reactions and self-discharge. These challenges of limited data and non-generalizability can be overcome with large transformers-based foundation models that have mastered extraction of molecular information in a self-supervised manner of learning from extensive unlabeled corpora. MoLFormer, a chemical large language model, has learned chemical and physical panorama of over 1 billion molecules from simple extractable chemical representations such as SMILES strings. As the result, MoLFormer’s reliance on labeled data for downstream task is significantly reduced. In present work, we demonstrate the predictive capabilities of a customized MolFormer model for battery electrolytes with a mere 140 electrolyte formulations datapoints. Gathering electrolyte formulations vs battery performance data from experimental process is a challenging endeavor. The proposed model predicts battery performance metrices such as coulombic efficiency and specific capacity based on 2 separate electrolyte formulation datasets with best accuracies among all the known machine learning models that simultaneously map structure and composition of formulation constituents. The potential of foundation models in designing mixed material systems such as liquid battery electrolytes present a groundbreaking opportunity to accelerate the discovery and optimization of new formulations across various industries.