Full-data accuracy with fewer labels for training and fine-tuning machine-learning force fields

Machine-learning force fields (MLFFs) are highly sensitive to their training data, often failing when applied to configurations outside their training distribution. While active learning (AL) can mitigate this by iteratively selecting informative data, standard uncertainty quantification (UQ) methods like model committees are computationally expensive, requiring multiple independent training runs. This paper introduces a workflow using last-layer-projection regression (LLPR) to estimate uncertainty in a single forward pass. The authors evaluate LLPR's ability to guide data selection across train-from-scratch and foundation-model fine-tuning regimes, comparing it against random sampling and traditional model committees.

LLPR provides a robust, quantitatively calibrated uncertainty signal that correlates strongly with actual prediction errors in energy and force. In train-from-scratch experiments on water/ice datasets, LLPR-driven data selection consistently outperforms random sampling, reaching the full-data accuracy ceiling with substantially fewer labels. Notably, LLPR achieves performance comparable to model committees—which require multiple training runs—using only a single model. In foundation-model fine-tuning, the LLPR signal effectively detects structural hallucinations and unphysical coordination in electrolyte systems, enabling automatic termination of the learning loop and ensuring model reliability without requiring an external oracle.

This work provides a scalable, computationally light UQ strategy that bridges the gap between training-set size and downstream MD reliability. By replacing expensive model committees with a single-pass LLPR estimator, researchers can significantly reduce the cost of generating high-quality training data for both custom and foundation MLFFs. The ability to detect deployment-time failures—such as structural hallucinations that conventional stability diagnostics miss—makes this approach particularly valuable for high-throughput materials discovery and complex electrolyte simulations.