Balancing Diversity and Efficiency in Training Datasets for Robust Machine Learning Potentials

Machine learning interatomic potentials (MLIPs) enable atomistic simulations at near first-principles accuracy. As model architectures mature, their reliability is increasingly constrained by training data quality. Here, we introduce ATLAS (Automated Training with Latent-space Aware Sampling), a unified and multifunctional framework based on the reversibly compressed latent space and online reliability verification. By co-designing diversity and efficiency criteria with a diversity-aware database generator and a manifold-aware active learning workflow, we steer exploitation toward sparse configurational regions while preventing premature convergence. Validated on metals, alloys, and metal oxides, ATLAS produces compact datasets exceeding foundation model training chemical spaces by orders of magnitude with up to 317 times fewer structures. ATLAS-MLIPs outperform universal MLIPs by up to 425 K in melting point accuracy for CuZn, while surface energy errors for IrO x drop from 1.21 to 0.04 J/m 2 . ATLAS provides a principled and transferable strategy for building robust MLIPs across diverse material classes.