Uncertainty Quantification in PINNs for Turbulent Flows: Bayesian Inference and Repulsive Ensembles

The practical pressure here is that turbulence closure models are notoriously underspecified: small changes in boundary conditions or model assumptions can produce wildly different flow predictions, so knowing how confident your neural network actually is matters as much as the prediction itself. Repulsive ensembles, which explicitly push ensemble members apart in parameter space to improve coverage, are the less-discussed piece of this work and the one most likely to influence how practitioners build production RANS solvers.

Uncertainty quantification is showing up across multiple fronts in recent coverage. The SegWithU paper from April 16 tackled the same core problem in medical imaging, using perturbation energy to estimate confidence in a single forward pass. That work and this one are converging on a shared concern: that predictions without calibrated confidence estimates are operationally incomplete, regardless of domain. The difference is deployment context. Medical segmentation tolerates post-hoc probes; turbulence simulation in engineering pipelines may require the Bayesian inference approach this paper pursues, where the physics constraints themselves shape the posterior.

Watch whether any of the three methods tested here, Hamiltonian Monte Carlo, dropout, or repulsive ensembles, shows consistent calibration advantage on benchmark RANS cases with known ground truth (channel flow, backward-facing step). If repulsive ensembles hold their edge there, expect them to displace standard deep ensembles in physics-constrained settings within two years.