The quality of machine-learning interatomic potentials (MLIPs) is fundamentally limited by the physical fidelity of their first-principles training data. Here, we reveal that unrestricted density functional theory (DFT) calculations are essential to maintain the performance of MLIPs, even for closed-shell materials. In this work, we chose $\mathrm{LiAlSi}{\mathrm{O}}_4$, a closed-shell system, as a representative example and generated two different training datasets by applying restricted and unrestricted DFT calculations to the same set of geometric configurations. The two trained types of GRACE MLIPs proposed different equilibrium states, while the unrestricted GRACE delivers lower force errors, better structural accuracy, and more reliable Li-ion transport descriptions. The difference becomes far more pronounced under extreme conditions: the unrestricted GRACE remains stable in higher temperature and irradiation simulations, whereas the restricted model shows unphysical Li clustering and rapid breakdown under collision cascades. These results suggest that spin-crossover is an important factor governing the transferability and robustness of MLIPs.

Close

PrevNext

View figure in article

• Spin state transition
• Deep potential molecular dynamics
• Density functional calculations
• High-throughput calculations