Machine learning-guided screening of phase-stable high-entropy Na cathodes to enable EVs and low-cost charging storage systems

Abstract

The large-scale electrification of transportation has increased the demand for cost-effective and durable stationary battery energy storage systems (BESSs), where conventional lithium-ion batteries suffer from long-term durability and low-temperature limitations. Sodium-ion batteries (SIBs) are promising alternatives for stationary applications due to their elemental abundance, low cost, and improved safety. Recent advances in machine learning (ML) have enabled data-driven battery materials design and improved understanding of structural instability and phase evolution; however, most ML-based studies have focused primarily on electrochemical property prediction rather than direct screening of phase stability. Among various cathode materials, O3-type layered sodium transition-metal oxides exhibit favorable electrochemical performance but suffer from severe phase instability, particularly O3–P3 phase transitions during cycling, which degrade cycle life and reliability. To address this challenge, we propose a high-throughput computational screening framework that integrates high-entropy doping strategies with ML to enhance phase stability in O3-type Na layered oxide cathodes. A total of 792 doped compositions were generated by combining five transition-metal elements. Formation energies were predicted using an ML model trained on the Materials Project database, and phase stability was quantified using the energy above hull. Subsequent density functional theory (DFT) calculations evaluated phase preference between O3 and P3 structures, volume changes during charge, and operating voltage windows relevant to stationary BESS operation, ultimately identifying five promising candidates. Furthermore, an active learning (AL)-based model was developed to directly predict O3–P3 phase preference, enabling efficient discrimination of phase-stable compositions without exhaustive DFT calculations. Overall, this study establishes an ML-assisted, thermodynamically grounded screening framework that moves beyond conventional property prediction toward phase-stable cathode discovery for long-cycle stationary energy storage.