Introduces e²IP, an equivariant evidential deep learning framework for ML interatomic potentials in molecular dynamics. Models atomic forces and uncertainties via 3x3 covariance tensors that rotate equivariantly. Outperforms ensembles in accuracy, efficiency, and data efficiency.
Key Points
- 1.Single-model UQ for aleatoric/epistemic uncertainty
- 2.Equivariant covariance for vector forces
- 3.Better balance on molecular benchmarks
Impact Analysis
Enhances reliability of MD simulations for materials science. Enables uncertainty-aware active learning and extrapolation detection. Reduces computational costs over ensemble methods.
Technical Details
Backbone-agnostic; fuses scalar evidential learning with equivariance. Symmetric positive definite tensors ensure self-consistency. arXiv:2602.10419v1.