Predictive Uncertainty for Neural CAE Surrogates

This research paper investigates how established uncertainty quantification (UQ) approaches behave when adapted to geometry-conditioned neural surrogates for computer-aided engineering (CAE) workflows. The authors compare one closed-form and two sampling-based UQ methods, evaluating their performance across varying geometries, spatial prediction fields, and different engineering quantities of interest. The study addresses the critical need for meaningful uncertainty estimates when using neural surrogates in design optimization scenarios.