Comparison of Machine Learning and Empirical Potentials for Vacancy Cluster Properties in Tungsten

Accurate modeling of vacancy clusters in tungsten is critical for advanced fusion reactor materials. Here, we systematically evaluate traditional empirical and machine learning potentials against density functional theory benchmarks. Using a Wigner–Seitz structure-energy framework, we reveal that traditional potentials systematically underestimate the surface energy penalty of microscopic voids. In contrast, machine learning potentials naturally capture the fundamental Wigner–Seitz scaling law without manual parameter tuning. Consequently, they accurately reproduce the geometric magic numbers and complex sawtooth fluctuations in binding energetics. Our work independently validates the analytical Wigner–Seitz scaling model and establishes machine learning potentials as a robust foundation for simulating vacancy-driven defect dynamics while emphasizing the necessity of property-specific benchmarking prior to large-scale or kinetic Monte Carlo simulations.