Mining Chemical Space with Generative Models for Battery Materials

Generative models represent a powerful new paradigm for accelerating the discovery of novel materials across vast chemical space. To evaluate the viability of deploying generalized crystal generative models for application specific discovery tasks, here we choose Li-ion battery(LIB) materials as a case study. The pretrained MatterGen model is used to generate diverse crystalline structures, conditioned for stability and tested for uniqueness and novelty for promising Li-containing compositions. We performed unsupervised clustering analysis using atomic neighborhood fingerprints to compare the distribution of generated structures against the training dataset and materials project(MP) data in the chemical space. Our multi-tired workflow for LIB materials combines universal crystal generative model with foundational machine learning potential to identify most promising stable (close to convex hull with respect to MP data) candidates for final Density Functional Theory (DFT) based stability calculations. Open Circuit Voltage (OCV) and specific capacity calculations on selected stable materials highlighted their potential as LIB materials. Among 91 identified Li-containing stable (0.03 eV/atom above MP convex hull) materials, we identified three novel cathode materials useful for LIB considering average OCV, OCV at highest and lowest state of charge, and the specific capacity.