Equivariant Evidential Deep Learning for Interatomic Potentials

Researchers developed e²IP, a new framework for uncertainty quantification in machine learning interatomic potentials used in molecular dynamics simulations. The method uses equivariant evidential deep learning to model atomic forces and their uncertainty through symmetric covariance tensors that transform properly under rotations.

• →e²IP addresses limitations of existing uncertainty quantification approaches for machine learning interatomic potentials by providing better computational efficiency.
• →The framework maintains statistical self-consistency under rotational transformations by using 3×3 symmetric positive definite covariance tensors.
• →Experiments show e²IP achieves superior accuracy-efficiency-reliability balance compared to ensemble methods and non-equivariant baselines.
• →The method enables uncertainty-aware workflows like active learning for training dataset construction in molecular dynamics simulations.
• →The backbone-agnostic framework retains single-model inference efficiency while improving data efficiency through equivariant architecture.