pith. sign in

arxiv: 2604.23300 · v1 · submitted 2026-04-25 · ⚛️ physics.flu-dyn

Bayesian neural network correction of RANS turbulence models with uncertainty quantification in separated flows

Tyler Buchanan , Ali Eidi , Richard P. Dwight This is my paper

Pith reviewed 2026-05-08 07:28 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords Bayesian neural networksRANS turbulence modelsuncertainty quantificationanisotropy correctionseparated flowsdata-driven modelingperiodic hill flowbackward facing step
0
0 comments X

The pith

Bayesian neural network corrections to anisotropy substantially improve RANS velocity predictions in separated flows

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes using a Bayesian neural network to learn corrections for Reynolds-averaged Navier-Stokes turbulence models in flows with separation. These corrections adjust the turbulent kinetic energy source term and add a tensor to account for anisotropy in the stresses. On the periodic hill case used for training, the anisotropy correction leads to much better velocity fields that capture separation and recirculation more accurately, while uncertainty estimates are reliable. When the same network is applied to a different flow, the curved backward-facing step, the velocity improvements remain but are less precise and the uncertainties do not cover the actual errors well enough. The work concludes that the main issues stem from the form of the corrections and how they interact nonlinearly in the solver, not from the neural network training process.

Core claim

A Bayesian neural network is trained to predict a turbulent kinetic energy source correction and a Reynolds stress anisotropy tensor correction from RANS input fields, using data from the periodic hill flow. Ensembles drawn from the posterior distribution of network weights are used to create multiple correction fields that are frozen and propagated through the RANS equations via Monte Carlo sampling. This yields corrected mean flow predictions that show substantial gains in accuracy for velocity, separation point, and recirculation when anisotropy is included, along with calibrated uncertainty intervals. Application to the unseen curved backward-facing step configuration demonstrates that 3

What carries the argument

Bayesian neural network that generates ensembles of k-source and anisotropy tensor corrections, propagated into the RANS solver by frozen-realization Monte Carlo

If this is right

  • A source term correction for turbulent kinetic energy alone matches the energy levels but has little effect on the mean velocity field
  • The tensorial anisotropy correction produces significant gains in the accuracy of predicted velocities, separation, and recirculation
  • These gains transfer qualitatively to an unseen flow configuration, although with lower precision and poor uncertainty calibration
  • The residual errors are traced to the restricted form of the corrections and nonlinear propagation in the solver rather than to the Bayesian neural network approximation

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Including additional flow configurations in the training set might improve the uncertainty estimates for new geometries
  • Developing correction formulations that better capture nonlinear interactions could address the observed under-coverage
  • This method could be extended to other types of turbulence models beyond RANS for broader uncertainty-aware simulations

Load-bearing premise

The dominant modeling errors in these separated flows can be represented by adjustments to the turbulent kinetic energy source and the anisotropy tensor, and that freezing these corrections during Monte Carlo propagation suffices to quantify the resulting uncertainty

What would settle it

Compare the ensemble of BNN-corrected velocity profiles and their uncertainty bands from the curved backward-facing step simulation directly to high-fidelity reference data to determine whether the true discrepancies are contained within the predicted uncertainty ranges at the stated confidence level

Figures

Figures reproduced from arXiv: 2604.23300 by Ali Eidi, Richard P. Dwight, Tyler Buchanan.

Figure 1
Figure 1. Figure 1: Fully Bayesian BNN architecture and variational layer detail. view at source ↗
Figure 2
Figure 2. Figure 2: ELBO loss computation. Both the coefficient and noise head weights contribute to the KL divergence term (with learnable α). To improve convergence, training is performed in two phases. First, a deterministic network, equivalent to a TBNN [15], with the same architecture is pretrained using a mean squared error objective, providing stable initialization for the posterior means of the Bayesian weights. Sec￾o… view at source ↗
Figure 3
Figure 3. Figure 3: MC inference with the fully Bayesian model. Each posterior weight sample produces a di view at source ↗
Figure 4
Figure 4. Figure 4: Training convergence during the Bayesian fine-tuning phase (40 000 epochs). Top-left: ELBO loss. Top-right: view at source ↗
Figure 5
Figure 5. Figure 5: Spatial fields on the periodic hill test set (RITA classified region only). (a) Reference view at source ↗
Figure 6
Figure 6. Figure 6: Spatial uncertainty decomposition on the periodic hill test set (RITA classified region only). (a) Epistemic view at source ↗
Figure 7
Figure 7. Figure 7: Variance explained (R 2 ) by projection of b ∆ i j onto combinations of the first four Pope basis tensors. The strain–rotation commutator (T2) dominates among single bases, while the combination T1, T2, T3 achieves near-complete representation of the anisotropy correction. Instead of predicting the full anisotropy tensor b ∆ i j directly, the model learns the scalar co￾efficients gn associated with the sel… view at source ↗
Figure 8
Figure 8. Figure 8: Spatial fields of the anisotropy correction components view at source ↗
Figure 9
Figure 9. Figure 9: Training convergence of the b ∆ i j BNN. 3.3. Stochastic propagation results The trained BNN corrections are propagated through the RANS solver as stochastic ensem￾bles. To isolate the physical role of each correction, propagation is performed in two stages: (i) using only kdeficit, and (ii) using the combined kdeficit + b ∆ i j model. Each MC sample is propa￾gated through the augmented k–ω SST solver (Ope… view at source ↗
Figure 10
Figure 10. Figure 10: shows vertical profiles of TKE at nine streamwise stations. As expected, baseline RANS systematically underpredicts the shear-layer TKE level throughout the separated region, with the largest discrepancies appearing around x/H ≈ 2–5. In contrast, the propagated BNN mean closely follows the BP reference at all stations, reproducing both the magnitude and loca￾tion of the TKE peak. The propagated uncertaint… view at source ↗
Figure 11
Figure 11. Figure 11: Vertical profiles of streamwise velocity view at source ↗
Figure 12
Figure 12. Figure 12: Contour comparison for TKE in the kdeficit-only propagation stage. (a) Propagated BNN mean k. (b) Difference between propagated BNN mean and BP reference. (c) Propagated epistemic uncertainty σ(k). The black dashed contour indicates the RITA-classified region where the correction is applied. 23 view at source ↗
Figure 13
Figure 13. Figure 13: Contour comparison for streamwise velocity in the view at source ↗
Figure 14
Figure 14. Figure 14: Observed coverage for the kdeficit-only propagation stage. The bars report the percentage of LES values falling within the propagated BNN ±1σ and ±2σ intervals for k and Ux. Gray dashed and dotted lines mark the nominal 95% and 68% Gaussian coverage levels, respectively. Overall, the kdeficit-only propagation stage demonstrates that the scalar BNN correction al￾ready captures a substantial part of the mis… view at source ↗
Figure 15
Figure 15. Figure 15: Vertical profiles of TKE k at nine streamwise stations for combined propagation. Baseline RANS (gray dashed) underpredicts the shear-layer peak. The BNN mean (green) closely follows the BP reference (red), while the ±2σ band captures the uncertainty. LES data are shown for comparison. The effect on the velocity field is substantially more pronounced, as shown in view at source ↗
Figure 16
Figure 16. Figure 16: Vertical profiles of streamwise velocity view at source ↗
Figure 17
Figure 17. Figure 17: Contour comparison for TKE in the combined propagation stage. (a) Propagated BNN mean view at source ↗
Figure 18
Figure 18. Figure 18: Contour comparison for streamwise velocity in the combined propagation stage. (a) Propagated BNN mean view at source ↗
Figure 19
Figure 19. Figure 19: Reconstructed anisotropy components b ∆ i j for combined propagation. Rows show BP reference, BNN mean, epistemic uncertainty, and difference (BNN − BP). Columns correspond to tensor components. The dashed contour indicates the RITA region. Coverage statistics for the combined propagation are shown in view at source ↗
Figure 20
Figure 20. Figure 20: Coverage statistics for combined propagation ( view at source ↗
Figure 21
Figure 21. Figure 21: Vertical profiles of turbulent kinetic energy view at source ↗
Figure 22
Figure 22. Figure 22: Vertical profiles of streamwise velocity view at source ↗
Figure 23
Figure 23. Figure 23: Contour comparison for turbulent kinetic energy in the CBFS case under view at source ↗
Figure 24
Figure 24. Figure 24: Contour comparison for streamwise velocity in the CBFS case under view at source ↗
Figure 25
Figure 25. Figure 25: Vertical profiles of turbulent kinetic energy view at source ↗
Figure 26
Figure 26. Figure 26: Vertical profiles of streamwise velocity view at source ↗
Figure 27
Figure 27. Figure 27: Predicted anisotropy correction fields for the CBFS case. Top row: BNN mean of view at source ↗
Figure 28
Figure 28. Figure 28: Contour comparison for turbulent kinetic energy in the CBFS case under combined propagation. (a) BNN view at source ↗
Figure 29
Figure 29. Figure 29: Contour comparison for streamwise velocity in the CBFS case under combined propagation. (a) BNN mean view at source ↗
read the original abstract

Data-driven correction of turbulence models offers a promising route for improving Reynolds-averaged Navier-Stokes (RANS) predictions, but quantifying uncertainty in such corrections and ensuring generalization across flows remain key challenges. This work presents a Bayesian neural network (BNN) framework for uncertainty-aware correction of RANS models. Two complementary correction mechanisms are considered: a turbulent kinetic energy source-term correction (k_deficit) and a tensorial anisotropy correction (b_ij^Delta). Posterior samples of the BNN weights are used to generate ensembles of deterministic correction fields, which are propagated through the RANS solver using a frozen-realization Monte Carlo approach. The framework is trained and evaluated on the periodic hill flow and further assessed on an unseen configuration, the curved backward-facing step. Results show that the k-source term correction alone accurately reproduces turbulent kinetic energy with well-calibrated uncertainty, but has negligible impact on the mean velocity field. In contrast, the inclusion of anisotropy correction leads to substantial improvements in velocity predictions, enabling more accurate representation of separation and recirculation. While these improvements persist qualitatively in the unseen case, reduced accuracy and significant under-coverage are observed, highlighting the challenges of out-of-distribution generalization and uncertainty quantification. Analysis of the results indicates that remaining discrepancies are primarily linked to limitations of the correction formulation and nonlinear propagation effects, rather than the BNN approximation itself. The proposed framework provides a physically consistent approach for propagating epistemic uncertainty in data-driven turbulence corrections and offers a robust pathway toward uncertainty-aware and generalizable RANS modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript develops a Bayesian neural network (BNN) framework to correct RANS turbulence models via two mechanisms—a turbulent kinetic energy source-term correction (k_deficit) and a tensorial anisotropy correction (b_ij^Delta)—with epistemic uncertainty quantified by propagating posterior weight samples through the RANS solver using frozen-realization Monte Carlo. Trained on periodic hill data and evaluated on an unseen curved backward-facing step, the work claims that anisotropy correction yields substantial improvements in mean velocity and separation predictions while k-correction alone reproduces TKE well; however, out-of-distribution accuracy drops and uncertainty under-coverage appears, which the abstract attributes primarily to correction-formulation limits and nonlinear propagation effects rather than the BNN itself.

Significance. If the central claims hold after addressing propagation issues, the framework offers a physically motivated route to embed epistemic uncertainty from data-driven corrections into RANS predictions for separated flows, with the separation of k-only versus anisotropy effects providing useful diagnostic insight. The approach builds on existing high-fidelity training data without introducing new ad-hoc constants, but the absence of quantitative error metrics or verified calibration in the provided abstract limits immediate assessment of impact.

major comments (3)
  1. [Abstract] Abstract and results discussion: the claim that anisotropy correction produces 'substantial improvements' in velocity and recirculation is presented qualitatively, yet no quantitative metrics (e.g., L2 velocity errors, separation-point displacement, or integrated recirculation strength) are reported to substantiate the magnitude or statistical significance of the gain over baseline RANS.
  2. [Methods / Results] Propagation method (described in methods and invoked in results): the frozen-realization Monte Carlo approach fixes each BNN-generated correction field before the RANS solve, thereby omitting iterative coupling in which updated k and anisotropy alter the mean flow, which in turn modifies the effective corrections. The abstract itself identifies 'nonlinear propagation effects' as a primary source of remaining discrepancies and notes significant under-coverage on the OOD case, directly implicating this approximation as load-bearing for both accuracy and UQ reliability.
  3. [Results (OOD case)] OOD evaluation: while qualitative persistence of improvement is asserted for the curved backward-facing step, the reported 'significant under-coverage' indicates that the posterior predictive intervals do not reliably contain the high-fidelity reference, undermining the uncertainty-quantification component of the central claim without further analysis of coverage probability or recalibration.
minor comments (2)
  1. [Introduction / Methods] Notation for the anisotropy correction tensor (b_ij^Delta) should be defined explicitly on first use and distinguished from the standard anisotropy tensor to avoid reader confusion.
  2. [Results] The manuscript would benefit from a table or figure summarizing quantitative error metrics (velocity, TKE, separation location) for baseline RANS, k-only correction, and full anisotropy correction on both training and test cases.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments and suggestions. We address each major comment below and indicate the revisions made to the manuscript.

read point-by-point responses

  1. Referee: Abstract and results discussion: the claim that anisotropy correction produces 'substantial improvements' in velocity and recirculation is presented qualitatively, yet no quantitative metrics (e.g., L2 velocity errors, separation-point displacement, or integrated recirculation strength) are reported to substantiate the magnitude or statistical significance of the gain over baseline RANS.

    Authors: We agree that quantitative metrics would strengthen the presentation of our results. In the revised manuscript we have added L2 velocity error norms, separation-point displacement, and integrated recirculation strength metrics (bubble area and peak reverse velocity) to both the abstract and results sections for the periodic hill case, with corresponding values also reported for the OOD curved backward-facing step to document the reduced accuracy. revision: yes

  2. Referee: Propagation method: the frozen-realization Monte Carlo approach fixes each BNN-generated correction field before the RANS solve, thereby omitting iterative coupling in which updated k and anisotropy alter the mean flow, which in turn modifies the effective corrections. The abstract itself identifies 'nonlinear propagation effects' as a primary source of remaining discrepancies and notes significant under-coverage on the OOD case, directly implicating this approximation.

    Authors: We acknowledge that the frozen-realization Monte Carlo method is an approximation that does not capture full iterative coupling between the corrections and the mean flow. This choice was made for computational tractability, as repeated full RANS solves per posterior sample would be prohibitive. We have expanded the Methods section to describe the approximation explicitly, its rationale, and its potential contribution to the nonlinear effects noted in the abstract. We have also added a limited sensitivity study on a subset of samples using partial iterative coupling to quantify the additional error introduced by the frozen approach. revision: partial

  3. Referee: OOD evaluation: while qualitative persistence of improvement is asserted for the curved backward-facing step, the reported 'significant under-coverage' indicates that the posterior predictive intervals do not reliably contain the high-fidelity reference, undermining the uncertainty-quantification component of the central claim without further analysis of coverage probability or recalibration.

    Authors: We agree that the under-coverage on the OOD case requires quantitative support. The revised manuscript now reports explicit coverage probabilities (fraction of reference points lying inside the 95% credible intervals) for both the training periodic hill and the OOD curved backward-facing step. We have also added discussion of possible recalibration approaches and reiterated that the under-coverage is attributed to correction-formulation limits and nonlinear propagation rather than the BNN itself, given the well-calibrated results on the training data. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper trains a BNN on independent external high-fidelity data (periodic hill) to learn k-source and anisotropy corrections, then applies posterior samples via frozen Monte Carlo to RANS solves on both training and OOD flows. The velocity improvements and UQ estimates are not equivalent to the inputs by construction; the framework explicitly discusses remaining discrepancies from correction formulation and nonlinear effects rather than claiming tautological reproduction. No self-definitional reductions, fitted inputs renamed as predictions, or load-bearing self-citations appear in the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard RANS modeling assumptions plus data-driven fitting of the BNN to high-fidelity data; no new physical entities are postulated.

free parameters (1)
  • BNN architecture and prior hyperparameters
    Chosen to fit correction fields to periodic-hill data; exact values not stated in abstract.
axioms (2)
  • domain assumption RANS equations remain a valid base model when supplemented by learned corrections
    Invoked throughout the correction framework description.
  • domain assumption Frozen-realization Monte Carlo adequately propagates epistemic uncertainty through the nonlinear solver
    Used to generate ensembles of corrected fields.

pith-pipeline@v0.9.0 · 5573 in / 1327 out tokens · 44825 ms · 2026-05-08T07:28:59.858819+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    H. Li, Y . Zhang, H. Chen, Aerodynamic prediction of iced airfoils based on modified three-equation turbulence model, AIAA Journal 58 (2020) 3863–3876

    work page 2020

  2. [2]

    H. Xiao, P. Cinnella, Quantification of model uncertainty in rans simulations: A review, Progress in Aerospace Sciences 108 (2019) 1–31

    work page 2019

  3. [3]

    A. Eidi, N. Zehtabiyan-Rezaie, R. Ghiassi, X. Yang, M. Abkar, Data-driven quantification of model-form uncer- tainty in reynolds-averaged simulations of wind farms, Physics of Fluids 34 (2022)

    work page 2022

  4. [4]

    Menter, Zonal two equation k-ωturbulence models for aerodynamic flows, 23rd fluid dynamics, plasmadynam- ics, and lasers conference (1993) 2906

    F. Menter, Zonal two equation k-ωturbulence models for aerodynamic flows, 23rd fluid dynamics, plasmadynam- ics, and lasers conference (1993) 2906

    work page 1993

  5. [5]

    D. C. Wilcox, et al., Turbulence modeling for CFD, volume 2, DCW industries La Canada, CA, 1998

    work page 1998

  6. [6]

    F. G. Schmitt, About boussinesq’s turbulent viscosity hypothesis: historical remarks and a direct evaluation of its validity, Comptes Rendus Mécanique 335 (2007) 617–627

    work page 2007

  7. [7]

    C. L. Rumsey, Exploring a Method for Improving Turbulent Separated-Flow Predictions with Kappa-Omega Mod- els, Technical Report NASA/TM-2009-215952, NASA Langley Research Center, Hampton, V A, USA, 2009. Tech- nical Memorandum

    work page 2009

  8. [8]

    Duraisamy, G

    K. Duraisamy, G. Iaccarino, H. Xiao, Turbulence modeling in the age of data, Annual Review of Fluid Mechanics 51 (2019) 357–377

    work page 2019

  9. [9]

    Ferson, L

    S. Ferson, L. R. Ginzburg, Different methods are needed to propagate ignorance and variability, Reliability Engineering & System Safety 54 (1996) 133–144

    work page 1996

  10. [10]

    Weatheritt, R

    J. Weatheritt, R. D. Sandberg, The development of algebraic stress models using a novel evolutionary algorithm, International Journal of Heat and Fluid Flow 68 (2017) 298–318

    work page 2017

  11. [11]

    Schmelzer, R

    M. Schmelzer, R. P. Dwight, P. Cinnella, Discovery of algebraic reynolds-stress models using sparse symbolic regression, Flow, Turbulence and Combustion 104 (2020) 579–603

    work page 2020

  12. [12]

    Pobe, Turbulent Flows, Cambridge University Press, 2005

    S. Pobe, Turbulent Flows, Cambridge University Press, 2005

    work page 2005

  13. [13]

    Mandler, B

    H. Mandler, B. Weigand, Generalization limits of data-driven turbulence models, Flow, Turbulence and Combus- tion (2024) 1–36

    work page 2024

  14. [14]

    Cherroud, X

    S. Cherroud, X. Merle, P. Cinnella, X. Gloerfelt, Sparse bayesian learning of explicit algebraic reynolds-stress models for turbulent separated flows, International Journal of Heat and Fluid Flow 98 (2022) 109047

    work page 2022

  15. [15]

    J. Ling, A. Kurzawski, J. Templeton, Reynolds averaged turbulence modelling using deep neural networks with embedded invariance, Journal of Fluid Mechanics 807 (2016) 155–166

    work page 2016

  16. [16]

    Geneva, N

    N. Geneva, N. Zabaras, Quantifying model form uncertainty in reynolds-averaged turbulence models with bayesian deep neural networks, Journal of Computational Physics 383 (2019) 125–147

    work page 2019

  17. [17]

    H. Tang, Y . Wang, T. Wang, L. Tian, Y . Qian, Data-driven reynolds-averaged turbulence modeling with gener- alizable non-linear correction and uncertainty quantification using bayesian deep learning, Physics of Fluids 35 (2023)

    work page 2023

  18. [18]

    G. Pash, M. Hassanaly, S. Yellapantula, a priori uncertainty quantification of reacting turbulence closure models using bayesian neural networks, Engineering Applications of Artificial Intelligence 141 (2025) 109821

    work page 2025

  19. [19]

    S. L. Brunton, B. R. Noack, P. Koumoutsakos, Machine learning for fluid mechanics, Annual review of fluid mechanics 52 (2020) 477–508

    work page 2020

  20. [20]

    C. Wu, S. Zhang, Y . Zhang, Development of a generalizable data-driven turbulence model: Conditioned field inversion and symbolic regression, AIAA Journal 63 (2025) 687–706

    work page 2025

  21. [21]

    Srivastava, C

    V . Srivastava, C. L. Rumsey, G. N. Coleman, L. Wang, On generalizably improving rans predictions of flow separation and reattachment, in: AIAA SCITECH 2024 Forum, 2024, p. 2520

    work page 2024

  22. [22]

    Buchanan, M

    T. Buchanan, M. L ˘ac˘atu¸ s, A. West, R. P. Dwight, Data-driven rans closures using a relative importance term analysis based classifier for 2d and 3d separated flows, Computers & Fluids (2025) 106899

    work page 2025

  23. [23]

    Magris, A

    M. Magris, A. Iosifidis, Bayesian learning for neural networks: an algorithmic survey, Artificial Intelligence Review 56 (2023) 11773–11823

    work page 2023

  24. [24]

    Kendall, Y

    A. Kendall, Y . Gal, What uncertainties do we need in bayesian deep learning for computer vision?, Advances in neural information processing systems 30 (2017)

    work page 2017

  25. [25]

    R. M. Neal, Bayesian learning for neural networks, volume 118, Springer Science & Business Media, 2012

    work page 2012

  26. [26]

    M. E. Tipping, Sparse bayesian learning and the relevance vector machine, Journal of machine learning research 1 (2001) 211–244

    work page 2001

  27. [27]

    A. Eidi, T. Buchanan, L. Jiang, R. P. Dwight, Physics-guided bayesian neural networks for zonal correc- tions and uncertainty quantification in separated flows, 2025. URL:https://arxiv.org/abs/2511.14534. arXiv:2511.14534

    work page arXiv 2025

  28. [28]

    Breuer, N

    M. Breuer, N. Peller, C. Rapp, M. Manhart, Flow over periodic hills–numerical and experimental study in a wide range of reynolds numbers, Computers & Fluids 38 (2009) 433–457. 47

    work page 2009

  29. [29]

    Bentaleb, S

    Y . Bentaleb, S. Lardeau, M. A. Leschziner, Large-eddy simulation of turbulent boundary layer separation from a rounded step, Journal of Turbulence (2012) N4

    work page 2012

  30. [30]

    J. L. Callaham, J. V . Koch, B. W. Brunton, J. N. Kutz, S. L. Brunton, Learning dominant physical processes with data-driven balance models, Nature communications 12 (2021) 1016. 48

    work page 2021