An Ensemble-Based, Active-Learning Approach to Refine Foundational MLIPs for Applications at Extreme Conditions

Machine-Learned Interatomic Potentials (MLIPs) have emerged as powerful and accurate tools for modeling complex materials, capable of reproducing quantum-chemical properties with near first-principles accuracy. However, their accuracy can degrade under extreme conditions, such as high temperatures, large deformations, or with high-energy defect states, where chemical properties are well-outside the training data. Refining MLIPs to model out-of-equilibrium effects remains challenging and can involve incorporating additional (higher-level) quantum chemical datasets or even directly fitting to experimental observables. Model retraining and fine-tuning require careful decisions depending on the MLIP architecture, available data, and hyperparameters, as well as efficient scheduling on high-performance computing resources. We introduce Ensemble-FF-Fit, a modular and extensible framework for fine-tuning classes of MLIPs such as MACE. Ensemble-FF-Fit employs ensemble variance analysis— across multiple fits of a single model and across distinct model classes— to actively guide the selection of new training structures, focusing sampling where uncertainty is highest. Coupled with our recently developed MatEnsemble backend for concurrent task execution, the framework enables user-defined objective functions that fit to experimental or synthetic microscopy data (XRD, 4D-STEM) in addition to traditional QM simulation data properties (energies, forces, stresses, etc.). We demonstrate that Ensemble-FF-Fit can accelerate fine-tuning of ML foundation models and improve their performance for far-from-equilibrium, phase-change processes such as recrystallization and beam-induced defect formation in transition metal dichalcogenides.