When Forces Disagree: A Data-Free Fast Diagnostic from Internal Consistency in Direct-Force Neural Network Potentials

ICLR 2026 Conference Submission25413 Authors

Keywords: NNIPs, Uncertainty, Pre-trained, Data-free, Physics-informed Uncertainty Estimate, Algorithmic Stability, Internal Consistency, Inter-head Influence, Multi-headed Architecture

TL;DR: We introduce a fast physics-informed uncertainty metric for pre-trained direct-force neural network potentials that leverages the model's internal physical inconsistency to achieve the data-free advantage of ensembles at the single-model speed.

Abstract: Direct-force neural network potentials (NNIPs) offer superior speed for atomistic simulations, but their reliability is limited by the lack of a fast and data-free uncertainty estimate to monitor the impact of non-conservativity and prediction errors. While ensembles are data-free but slow, and other single-model methods often require training data, we introduce an approach that combines the advantages of both. Our metric is derived from the internal disagreement between a model's directly predicted force and its energy-gradient-derived force, motivated by our finding that a model's internal self-consistency is more critical for algorithmic stability than its external accuracy. We then identify an asymmetric failure mode inherent to the direct-force architecture that this metric can diagnose, and also show a strong monotonic correlation between the disagreement and the true force error across diverse materials and out-of-distribution structures. We propose the link between internal disagreement and practical reliability is a consequence of inter-head influence via the shared graph neural network embedding. We provide direct evidence for this mechanism by showing that fine-tuning the conservative force pathway on adversarial data that maximizes this internal disagreement measurably improves the stability of simulations driven only by the direct force. The metric serves as a versatile and out-of-the-box tool that is competitive with expensive ensembles, offering both an on-the-fly assessment of model reliability and a principled method for generating targeted data to improve the stability of direct-force models.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 25413