Answers are grounded in this paper and cite their sources.
Machine learning interatomic potentials (MLIPs) offer near–ab initio accuracy at a fraction of the cost of conventional electronic structure calculations, but their reliability depends critically on the quality and diversity of training data. Here, we present a classical-to-density functional theory (DFT) multi-fidelity workflow for efficient dataset generation, in which long classical molecular dynamics trajectories are supplemented with selectively DFT-optimized configurations to train neural network potentials. We demonstrate this approach on four systems: bulk water, water at the air–water interface, CaSO4 in bulk water, and CaSO4 at the interface. In all cases, the trained models faithfully reproduce reference DFT properties, validating the workflow across both simple and chemically complex environments. We further probe the extrapolation capabilities of the models, showing that bulk-trained networks can generalize to different temperatures, pressures, and interfacial geometries when datasets include targeted augmented configurations. This multi-fidelity strategy thus enables accurate and transferable MLIPs for complex solvated and interfacial systems, bridging the gap between computational efficiency and first-principles accuracy.
No takes yet. Share an insight, caveat, or question.
Siddique et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: