XRDMatch: a semi-supervised learning framework to efficiently discover room temperature lithium superionic conductors

Abstract

The long-sought prediction pipelines for solid-state electrolytes (SSEs) with room-temperature superionic conductivity mark a significant milestone on the path towards realizing the commercialization of all-solid-state lithium batteries. In recent years, machine learning (ML) has shown significant promise in accelerating the discovery of new materials, optimizing manufacturing processes, and predicting battery cycle life. However, material datasets are often smaller (with just a few hundred lithium-ion conductors) and, at times, more diverse, posing the challenge of training a reliable model as a key obstacle in accelerating material discovery. In response to this challenge, we pioneeringly proposed a semi-supervised learning framework integrating consistency regularization and pseudo-labeling, which only uses an X-ray diffraction (XRD) pattern as a descriptor without human intervention, named ‘XRDMatch’. Leveraging a wealth of unlabeled data information from the Inorganic Crystal Structure Database (ICSD) database to support limited labeled data, our approach aids in constructing accurate and robust models, with an F1 score of the ensemble learning strategy model reaching as high as 0.92. Further predictions on unlabeled data identify 38 superionic conductors, including 32 validated by recent literature reports and six new candidates quantified through ab initio molecular simulation. Among these, Li₆AsSe₅I was further synthesized and experimentally confirmed as a superionic conductor. This work underscores the feasibility of a semi-supervised learning framework in overcoming constraints posed by limited data and highlights the model's promising potential for efficiently discovering room-temperature superionic conductors.