关键词: 机器学习原子势, 层状正极材料, 电化学性能, 分子动力学, 阳离子混排

Abstract: Layered LiNi_xMn_1–xO₂ cathode materials undergo complex structural evolution during charge–discharge processes, including lithium-ion extraction/insertion, transition-metal migration, and cation mixing. The large configurational space of these materials makes conventional first-principles simulations computationally expensive. Machine learning interatomic potentials (MLIPs), which combine the accuracy of first-principles calculations with the efficiency of molecular dynamics simulations, provide an effective approach for exploring microscopic structural evolution. However, constructing reliable MLIPs for multicomponent Li–Ni–Mn–O systems with diverse local chemical environments remains challenging when using limited training data. This study proposes a machine learning potential development strategy that integrates classical molecular dynamics sampling with density functional theory (DFT) calibration. Multiple-temperature molecular dynamics simulations were performed to explore the structural configuration space of LiNi_xMn_1–xO₂, followed by density-based clustering to select representative configurations. These configurations were then combined with DFT data to train and calibrate an equivariant message-passing neural network based on the MACE framework, enabling the construction of an accurate MLIP for Li–Ni–Mn–O systems. The developed potential exhibits high predictive accuracy, with mean absolute errors (MAEs) of 0.0082 eV/atom for energy and 0.0693 eV/Å for atomic forces on an independent test set. The calculated delithiation voltage profiles reproduce experimental trends, demonstrating the capability of the model to capture the energetic characteristics and structural responses of LiNi_xMn_1–xO₂. Furthermore, combined with ground-state structure searches, the evolution of cation mixing under different Ni/Mn ratios and delithiation states was systematically investigated.The results reveal that structural instability in Ni-rich compositions mainly originates from enhanced Li-Ni antisite mixing during delithiation. The layered structure stability exhibits a distinct "dual-critical" behavior. A critical Mn content threshold of ~0.5 exists, below which the "structural pillar" effect of Mn⁴⁺ is insufficient to suppress Ni migration into the Li layer, while above which Jahn-Teller distortion of Mn³⁺ instead triggers severe Li-Mn mixing. A critical delithiation degree of 0.5–0.6 exists, beyond which extensive oxidation of Ni to high-valence states (Ni³⁺/Ni⁴⁺) causes ionic radius contraction and abrupt lattice stress accumulation, leading to a sharp surge in the mixing degree. Among the compositions studied, LiNi_0.4Mn_0.6O₂ exhibits superior resistance to cation mixing. This work provides a reliable computational methodology for simulating structural evolution and optimizing the composition design of complex cathode materials

Key words: machine learning interatomic potential, layered cathode materials, electrochemical performance, molecular dynamics, cation mixing