MLIP-DFT Combined High-Throughput Screening of Alkaline Stable Lithium Conductors

Solid-state Li-air batteries can prevent the parasitic reaction between the highly reactive discharge product Li 2 O 2 and the organic electrolytes, which is a common challenge for the conventional dry Li-O 2 battery. Using humidified air can further improve the battery performance by converting the discharge product to LiOH –– a less reactive and more reversible product. To achieve such solid-state humidified Li-air batteries, it is highly desirable to design Li ion conductors that can withstand humid and highly alkaline environment established by the highly saturated LiOH solution on the material surface. In this work, we combine the pre-trained universal machine learning interatomic potential (MLIP) –– Crystal Hamiltonian Graph neural Network (CHGNet) –– and the density functional theory (DFT) calculations to conduct a large-scale hierarchical high-throughput computational screening of Li conductors. Our screening strategy is to choose a known Li superionic conductor framework and to optimize the compositions by examining the material stabilities targeting for the real battery operating conditions. We developed specific design criteria to achieve better stability in alkaline environments. CHGNet structural relaxations are performed in the first screening stage to evaluate the energy of several hundred thousand different compositions. These candidate compositions are screened based on the various stability criteria. The structures of candidates passed the CHGNet screening (stage-1 candidates) are further relaxed with DFT to obtain energies with higher accuracy. The candidates are re-evaluated with stricter stability conditions, leading to the identification of lithium conductors characterized by both high synthesizability and robust stability against the operating conditions of humidified Li-air battery (stage-2 candidates). Our screening campaign demonstrates that owing to the recent development of universal MLIP with near-quantum accuracy, it is now possible to conduct computational evaluation of compositions with orders of magnitude larger chemical space than what conventional ab-initio high-throughput screening was capable of. Our final candidates not only include some of the known lithium superionic conductors, but also the new compositions with high alkaline stability, indicating the effectiveness of using MLIP for pre-screening purposes. Our strategy can be easily transferred to different material frameworks or other materials applications. We anticipate that with a continuous improvement of MLIPs, the scalability of in-silico material design can be significantly enhanced.