Bayesian neural operators miss systematic error: residual-informed correction cuts thermal-break online optimization from 178 hours to 18 minutes
Synopsis
Bayesian neural operators treat weight uncertainty as the only source of doubt, yet a learned operator keeps a systematic gap to the PDE solution that does not vanish as training data grow. The residual-informed Bayesian neural operator (RI-BNO) turns a one-step Newton correction of the governing residual into error observations, fits a probabilistic inadequacy model, and folds its mean and covariance into the likelihood. On a reaction-diffusion benchmark, glioma MRI inversion and 3D thermal-break risk optimization it agrees better with finite-element references than the vanilla Bayesian neural operator while using half as many full nonlinear or transient training simulations; thermal-break online optimization falls from about 178 hours to 18 minutes.
Interpretation
Neural-operator prediction error is not only weight uncertainty; it also carries systematic bias from finite-dimensional representations, architectural restrictions and imperfect training, and that bias leaves an accuracy plateau as the training set grows. Earlier Bayesian neural operators modeled only weight uncertainty from finite training data, so their predictive distributions stayed conditional on the chosen representation, architecture and inference approximation and did not explicitly represent structured discrepancy from the PDE solution operator. On the nonlinear reaction-diffusion benchmark, the vanilla Bayesian neural operator approached architecture-dependent accuracy ceilings for both the latent-subspace and DeepONet architectures as the observation budget grew, while RI-BNO changed that scaling behavior and reached higher accuracy at a smaller total observation budget.
RI-BNO converts the governing PDE residual into extra observations of operator error, then learns an input-dependent probabilistic inadequacy model whose mean gives the systematic correction and whose covariance gives the uncertainty in that correction. The residual observations come from a one-step Newton correction linearized at the neural operator prediction, so no full nonlinear finite-element solution is generated at that input; a sparse variational Gaussian process fits the inadequacy model, which is then held fixed and used to build an inadequacy-aware likelihood for re-inferring the weight posterior. In glioma MRI inversion with the same pretrained operators and no patient-specific retraining, RI-BNO reconstruction reached a DICE of 0.918 against 0.724 for the vanilla operator and 0.932 for finite elements; across four additional BraTS 2019 patients, RI-BNO averaged 0.912 against 0.732 and 0.936.
Carrying operator inadequacy into downstream inference and design changes not only forward accuracy but the inferred parameter fields, tail-risk estimates and the design finally selected. The vanilla Bayesian neural operator's forward bias propagated into the inverse problem and the optimization: in the thermal-break design it produced an over-extensive fiber network that pushed the response into the insulation-failure regime. In CVaR risk optimization of a 3D thermal break, RI-BNO recovered a fiber distribution and tail risks close to the finite-element design, cut online optimization from about 178 hours with finite elements to 18 minutes, and lowered end-to-end cost by 17.4% relative to the vanilla Bayesian neural operator.
Perspective
The demonstrated scope is one nonlinear reaction-diffusion benchmark, one axial T1-CE glioma MRI inversion from BraTS 2019 plus reuse on four further patients, and CVaR risk optimization of one 3D fiber-reinforced aerogel thermal break; every comparison is against finite-element solutions rather than clinical or manufacturing measurements.
The local error bound for the residual observations rests on stability and smoothness assumptions and covers the estimator before inadequacy learning and full Bayesian inference; the authors state it is not a global guarantee or a calibration result for the final posterior, and the Gaussian-process inadequacy model with modular inference approximates the dependence between weights and the inadequacy field.
