Geometry-Aware Post-Hoc Uncertainty Quantification in Operator Learning
Researchers have introduced REEF-GP, a novel post-hoc uncertainty quantification framework for neural operators. This method fits a Gaussian Process to the residuals of a frozen neural operator, leveraging its internal embeddings to create geometry-aware uncertainty estimates. REEF-GP incorporates spectral-normalized projections and efficient subset-based training to ensure stability and scalability, outperforming deep ensembles in calibration and cost across various PDE benchmarks while remaining robust to geometric distribution shifts. AI
IMPACT Enhances the reliability of neural operators for complex scientific simulations by providing calibrated uncertainty estimates.