Structure-Aware Epistemic Uncertainty Quantification for Neural Operator PDE Surrogates

A research team has introduced a novel method for quantifying the 'epistemic' uncertainty—the uncertainty stemming from limited data and imperfect models—in Neural Operators (NOs). NOs are AI models that act as fast, resolution-invariant surrogates for solving complex Partial Differential Equations (PDEs), crucial for simulations in fields like fluid dynamics and material science. However, their predictions can be unreliable under data scarcity or novel conditions. The new 'structure-aware' technique addresses this by strategically injecting uncertainty only into the network's initial 'lifting' module, which encodes the input field, rather than applying unstructured perturbations like standard dropout across the entire model. This respects the modular 'lifting-propagation-recovery' architecture of modern NOs, treating the learned solver dynamics as deterministic.

This targeted approach results in uncertainty estimates that are both more computationally efficient and spatially faithful. Experiments on challenging benchmarks, including discontinuous-coefficient Darcy flow and geometry-shifted 3D computational fluid dynamics (CFD) simulations for car aerodynamics, demonstrate significant improvements. Compared to common baselines, the method provides more reliable statistical coverage of the true error, produces tighter and more useful prediction bands, and achieves better alignment between the estimated uncertainty and the actual localized residual structures in the simulation. For engineers and scientists using AI surrogates for high-stakes design or risk analysis, this means a more trustworthy and actionable understanding of when and where the model's predictions might be wrong.