Hiccup: A Resource-Efficient Automated Platform for High-Performance Global Neural Network Potentials Training

Abstract Neural network potentials (NNPs) enable atomistic simulations with near-first-principles accuracy at greatly reduced cost, yet the construction of specialized high-accuracy NNPs for complex systems remains limited by expert-dependent workflows and the challenge of achieving high accuracy and robust generalization under limited computational budgets. Here we present Hiccup, a resource-efficient automated platform for training high-performance global NNPs. Hiccup leverages prior guidance from universal NNPs to overcome the cold-start bottleneck of active learning, and integrates this initialization with hierarchical genetic algorithm–molecular dynamics (GA–MD) sampling to enable both broad configurational exploration and targeted refinement of low-energy basins. Building on these algorithmic advances, the platform further implements an asynchronous submit-and-forget scheduling mechanism that minimizes resource idling by overlapping first-principles labeling with subsequent sampling, training, and evaluation on the GPU. We demonstrate the application and efficiency of Hiccup using the Cu–Y–O ternary system as a case study. The resulting system-specific NNP achieves near-first-principles accuracy in static predictions and exhibits robust long-time dynamical fidelity across a broad thermodynamic range. These results establish Hiccup as a practical and scalable infrastructure for automated construction of high-quality interatomic potentials for chemically complex systems.