Efficient Evidential Deep Learning for Atomic Potentials
Original Title: Evidential deep learning for interatomic potentials
Journal: Nature communications
DOI: 10.1038/s41467-025-67663-y
Overview
Molecular dynamics simulations are essential for understanding physical and chemical processes at the atomic level. Machine learning interatomic potentials have emerged as an efficient tool to achieve the accuracy of first-principles calculations with significantly reduced computational costs. However, the reliability of these models depends on the quality of training data. Uncertainty quantification is required to identify when a model encounters unfamiliar atomic configurations. Conventional methods like ensemble modeling require training multiple networks, which is computationally expensive and slow. This paper introduces the evidential interatomic potential, a framework that provides robust uncertainty estimates through a single forward pass. By integrating evidential deep learning with a physics-inspired architecture, the model achieves high efficiency without sacrificing predictive accuracy.
Novelty
The novelty of this work lies in three specific design choices tailored for atomic systems. First, it implements locality by estimating uncertainty at the atomic level rather than for the entire system, utilizing atomic forces as a proxy for ground truth. Second, it accounts for directionality by producing separate uncertainty values for each Cartesian component of the force, recognizing that model confidence can vary across different spatial axes. Third, the framework replaces standard Gaussian assumptions with a Bayesian quantile regression approach. This allows the model to learn the distribution of forces more flexibly. In benchmarks on the ISO17 dataset, the method demonstrated strong performance, with Spearman’s rank correlation coefficients between 0.74 and 0.86 and ROC-AUC values ranging from 0.86 to 0.93. On a universal potential trained on millions of configurations, the model achieved an ROC-AUC of 0.914.
Potential Clinical / Research Applications
This framework facilitates more efficient active learning workflows, where the model can autonomously identify and sample high-uncertainty configurations to improve itself. This is particularly useful in drug discovery and materials design, where exploring the vast chemical space is otherwise prohibitive. Furthermore, the authors demonstrate uncertainty-driven dynamics, where the potential energy surface is modified to encourage the exploration of diverse atomic arrangements. In tests on water and lithium iron phosphate, this approach generated more diverse configurations than standard molecular dynamics. Such capabilities are vital for studying rare events and complex biochemical reactions. The ability to perform real-time assessment of model reliability during simulation ensures that researchers can trust results covering 89 different elements without the need for constant, expensive manual verification against first-principles methods.