Automated Workflow Toolkit for Training Reactive Machine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs) have become essential for extending the reach of atomistic simulations for larger system and longer time scales, enabling the study of complex chemical processes with near ab initio accuracy. However, constructing reactive and transferable MLIPs remains challenging, due to the need for high quality training datasets that include rare-events, high energy intermediates and transition states. In this work we present SPARC (Smart Potential with Rare Event and Continuous Learning), a modular Python workflow package design to automate the construction of reactive MLIPs and to generate near accurate potential energy surface (PES) model. SPARC implements an active learning loop that couples advance sampling techniques to systematically identify and label chemically relevant configurations. The workflow consists of three core steps, accelerated sampling driven PES exploration, labelling using VASP, CP2K, and model training with DeePMD-kit framework. All three steps are executed in an iterative fashion with minimal human intervention. A flexible YAML-based input configuration supports checkpoint, restart, and dynamic data management. In addition to that, it also provides utilities for visualizing new configurations, monitoring uncertainty in forces, and checking model performance across the iterations. We demonstrate the utility of SPARC on small molecular systems, showcasing its ability to construct robust MLIPs through automated exploration of complex chemical phase space.