Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials

The development of protective coatings for solid-state batteries is critical for improving the safety and performance of next-generation energy storage systems. These coatings must be electrically insulating and electrochemically stable in contact with both the solid electrolyte and electrode materials, while also enabling efficient Li-ion transport. Traditionally, candidate materials are identified by screening existing materials databases. However, these databases are approaching saturation, which limits the likelihood of discovering fundamentally new compounds. While generative artificial intelligence offers an alternative approach by enabling the design of previously unexplored materials, the vast size of the crystal-chemical search space makes systematic exploration computationally demanding, requiring efficient high-throughput workflows. We introduce a computational pipeline for discovering protective coating materials for Li-ion solid-state batteries, built on universal machine learning interatomic potentials (uMLIPs) that play the role of foundation models in atomistic simulation. The pipeline integrates three stages: (i) generation of candidate crystal structures using WyFormer, a symmetry-aware transformer based on Wyckoff representations; (ii) high-throughput evaluation of thermodynamic stability and electrochemical compatibility with the prominent solid electrolyte and electrode materials via uMLIP-driven structure optimization; and (iii) screening of Li-ion transport using descriptors derived from uMLIP-predicted potential energy surfaces. This descriptor-based approach enables identification of promising candidates in chemical spaces not represented in existing datasets. Top candidates are further validated using machine learning-accelerated molecular dynamics, to be followed by higher-accuracy ab initio calculations. The pipeline is modular, enabling scalable execution across heterogeneous high-performance computing resources. This framework provides a practical approach for exploring large chemical spaces and discovering functional materials by combining generative models with foundational atomistic potentials.