Property-balancing active learning strategy for machine learning interatomic potentials: case study on dislocation mechanics in magnesium

Interatomic potentials remain a critical bottleneck for accurate large-scale atomistic simulations. Machine-learning potentials provide a promising route to overcoming this limitation; however, the construction of property-relevant training datasets remains a major challenge. Here, we propose an active-learning strategy based on property balancing, in which the fraction of atomic environments associated with pre-identified key material properties is dynamically optimized during training. We demonstrate this strategy for pure magnesium, whose macroscopic plastic deformation is governed by several subtle dislocation transformations that place stringent demands on potential accuracy. Using the proposed strategy, we develop a Behler–Parrinello-type neural-network potential that captures these dislocation-mediated mechanisms with a level of accuracy that has been difficult to achieve in previous studies of magnesium. The successful application to magnesium highlights the versatility and effectiveness of property balancing as a general strategy for developing machine-learning potentials for metallic systems with multiple competing property requirements. This approach provides a practical pathway toward predictive mechanistic simulations of structural metals.