Advanced Machine Learning Interatomic Potential for Accelerated Atomistic Simulations of Lithiation Dynamics in Large-Scale Si@C Core-Shell Anodes.

Designing high-capacity silicon-based anodes for lithium-ion batteries is fundamentally challenged by severe volume expansion during lithiation, which limits cycle life and compromises structural stability. This challenge underscores the importance of achieving atomic-scale insight into lithiation dynamics. Herein, we report the development of a highly efficient neuroevolution potential (NEP) model that enables atomistic simulations of lithiation in large-scale silicon-carbon core-shell anodes, achieving a computational speed ∼70,000 times higher than conventional ab initio molecular dynamics while maintaining near-density functional theory accuracy. A direct sampling strategy was employed to reduce the training data set to 4.3% of the original configuration space, ensuring computational efficiency while preserving broad structural diversity. The validated NEP model accurately reproduces atomic forces, radial distribution functions, and lithium diffusivities across diverse structural motifs, including crystalline, amorphous, interfacial, and highly strained configurations. Systematic large-scale simulations reveal that, within the examined shell thickness range, a ∼4 nm carbon layer most effectively suppresses overall volume expansion (<1%) and establishes an inner-expansion/outer-locking mechanism that governs lithium distribution and mechanical confinement. These results provide fundamental atomic-scale insight into mitigating lithiation-induced volume change. Overall, this work demonstrates the substantial capability of machine-learning interatomic potentials to bridge the gap between computational feasibility and atomistic accuracy, offering a powerful framework for the rational design and optimization of next-generation silicon-carbon anodes with enhanced cycling stability and performance.