Screening “Cathode-Stable and Anode-Reactive” Sulfur-Based Additives for LiNi0.5Mn1.5O4||Graphite Pouch Cells via a Machine Learning Approach

Abstract For high-voltage spinel LiNi0.5Mn1.5O4 (LNMO) systems, under elevated temperature and high-voltage conditions, electrolyte decomposition and the associated interfacial “cross-talk” between the cathode and anode severely compromise cycling stability. To address this challenge, we present a machine learning (ML)-based approach with a gradient boosting regression (GBR) model to predict the highest occupied molecular orbital (HOMO) energy levels, which identifies the S═O group as a key structural element influencing additive effectiveness. Subsequently, by integrating density functional theory (DFT) calculations with molecular descriptor analysis, BDTT is identified as a promising candidate. Electrochemical evaluations demonstrate that, at 4.85 V and 45 °C, the incorporation of BDTT significantly enhances the capacity retention of LNMO||artificial graphite (AG) pouch cells from 24.71% to 82.73%. Multiscale characterizations reveal that BDTT undergoes preferential reduction at the anode, contributing to the formation of a sulfur-rich and compact solid electrolyte interphase (SEI), which effectively lowers interfacial resistance and suppresses cross-talk effects. This work validates the effectiveness of machine learning in functional additive discovery and provides new insights into electrolyte design for high-voltage lithium-ion batteries.