When Forces Disagree: A Data-Free Fast Uncertainty Estimate for Direct-Force Pre-trained Neural Network Potentials

Chayaphol Lortaraprasert, Rafael Gomez-Bombarelli

Keywords: NNIPs, Uncertainty, Pre-trained, Data-free, Physics-informed Uncertainty Estimate

TL;DR: We introduce a fast physics-informed uncertainty metric for pre-trained 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:

Neural Network Interatomic Potentials (NNIPs) are a cornerstone of modern atomistic simulations, but their reliability is limited by the difficulty in quantifying prediction uncertainty.
Current uncertainty quantification (UQ) methods present a trade-off: model ensembles offer a robust, data-free metric based on model disagreement but are computationally expensive, while faster single-model methods typically require access to the original training data which can be practically inconvenient and chemically sparse.
This paper introduces a novel differentiable UQ metric for direct-force pre-trained models that combines the strengths of both paradigms, offering the data-free reliability of ensembles with the computational speed of a single model.
Our metric is derived from the internal disagreement between two force predictions from a single NNIP—the directly predicted (non-conservative) force and the energy-gradient-derived (conservative) force.
We show a strong monotonic correlation between this force disagreement and the true force error against Density Functional Theory calculations.
This relationship is robust across a diverse set of materials and holds even for out-of-distribution structures generated via adversarial attacks.
Because the method is computationally cheap and requires no training data, it offers a powerful, out-of-the-box tool for on-the-fly assessment of model confidence with wide-ranging applications for reliable atomistic modeling.

Submission Track: Paper Track (Short Paper)

Submission Category: AI-Guided Design

Institution Location: Cambridge, USA

AI4Mat RLSF: Yes

Submission Number: 107