Physics-guided Bayesian neural networks for zonal corrections and uncertainty quantification in separated flows

Ali Eidi, Tyler Buchanan, Letian Jiang, Richard P. Dwight · 2025

Read Physics-guided Bayesian neural networks for zonal corrections and uncertainty quantification in separated flows by Ali Eidi, Tyler Buchanan, Letian Jiang, Richard P. Dwight (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-14534.txt.

Introduction

Ali Eidi and three co-authors present a Bayesian neural network method to correct turbulence models for separated flows. Standard averaged flow models lose accuracy where flow detaches from a surface. Data driven fixes often fail on new cases and give no reliable uncertainty. The team focus training on zones near separation to target the core problem.

Physics guided inputs keep the network consistent with known turbulence rules. A classifier applies corrections only where they are needed. Uncertainty from the Bayesian network flows through to the final flow solution. That lets users judge trust on unseen cases. The framework aims at more reliable engineering predictions for separated flow regimes.

Catalog record

CREATORAli Eidi, Tyler Buchanan, Letian Jiang, Richard P. Dwight
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-14534

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-guided Bayesian neural networks for zonal corrections and uncertainty quantification in separated flows Ali Eidi, Tyler Buchanan, Letian Jiang, Richard P. Dwight Abstract Data-driven techniques have improved the accuracy of Reynolds-averaged Navier-Stokes (RANS) models in fluid dynamics. However, modeling separated flows remains challenging due to their complex physics and sensitivity to local conditions. Existing approaches often struggle with generalization beyond training cases and lack robust uncertainty quantification frameworks, limiting their utility in complex flow regimes. We propose a Bayesian neural network (BNN)-based framework specifically designed for two-dimensional separated flows. By focusing on flow zones near separated regions, we ensure targeted training and enhance predictive reliability. The BNN framework incorporates physics-guided, invariant inputs to maintain consistency with turbulence physics. Correction terms predicted by the BNN are selectively applied to specific regions of the flow domain using a novel classifier, improving accuracy. A key feature of this approach is propagating BNN-derived corrections to flow solutions, enabling uncertainty quantification in unseen test cases. This probabilistic characterization of modeling errors offers insights into the reliability of RANS predictions across geometries with similar topologies. Preliminary results demonstrate that this method accurately predicts correction terms for Reynolds stress anisotropy and turbulent kinetic energy production in separated flow regions, effectively addressing dominant modeling errors and advancing turbulence modeling through uncertainty quantification. I Introduction Reynolds-averaged Navier–Stokes (RANS) simulations are widely used for turbulence modeling due to their efficiency but struggle in separated flows, where empirically tuned closures fail to capture complex flow physics. While large-eddy simulation (LES) improves accuracy, its high computational cost limits practical applications, driving the need for enhanced RANS models that balance accuracy and efficiency. Data-driven methods have been explored to refine RANS predictions, from parameter tuning to machine-learning-based Reynolds-stress corrections. Notably, symbolic regression 9 provides interpretable corrections but is constrained by predefined functional forms, while tensor-basis neural networks 5 offer greater flexibility but lack interpretability and are challenging to integrate into CFD solvers. Moreover, RANS models and most of the existing data-driven approaches provide deterministic corrections without quantifying uncertainty, which is crucial for reliable turbulence predictions. They often overlook epistemic uncertainty, which arises from model limitations and can be reduced with more data, and aleatoric uncertainty, which stems from inherent variability in the data and remains even with additional observations. Accounting for these uncertainties enhances the robustness of RANS simulations, especially in separated flows, where turbulence behavior is highly sensitive to local conditions. To address these challenges, this study proposes a Bayesian neural network (BNN)-based framework for RANS turbulence model corrections, extending BNNs beyond uncertainty quantification 7 to correction terms in separated flows. A tensor-based BNN estimates corrections with associated uncertainties, guided by a classifier that isolates shear-dominated regions for targeted input selection. This approach ensures physically consistent corrections, avoids unnecessary modifications, and propagates uncertainties into flow solutions, enhancing RANS prediction reliability across different geometries. The remainder of this paper is structured as follows: Section II presents the methodology, detailing the turbulence model corrections, the BNN formulation, and the classifier strategy. Section III discusses the results, focusing on production and Reynolds-stress corrections. Section IV summarizes the key findings and outlines potential future research directions. II Methodology II.1 Bayesian neural networks BNNs extend standard neural networks by placing probability distributions over their weights, enabling uncertainty quantification and incorporation of prior knowledge. Instead of optimizing a single set of weights, BNNs infer a posterior distribution using Bayes’ theorem, where the posterior over the weights p(w\mid D) is proportional to the product of the likelihood p(D\mid w) and the prior p(w) . Since direct computation is intractable, variational inference approximates the posterior by minimizing the Kullback–Leibler (KL) divergence between an approximate and true posterior, \mathrm{KL}\!\left(q(w\mid\theta)\,\|\,p(w\mid D)\right)=\int q(w\mid\theta)\,\log\!\left(\frac{q(w\mid\theta)}{p(w\mid D)}\right)\,dw, (1) as shown in Eq. ( 1 ). The choice of prior significantly impacts model performance by constraining weight distributions, influencing generalization and uncertainty estimation. While isotropic Gaussian priors are common, alternative approaches such as warm-start methods, Gaussian process initialization, and trainable priors offer improved flexibility 7 . This study employs a trainable prior, similar to KL annealing, where the KL term is dynamically adjusted to balance prior influence and data fitting. To make optimization tractable, the KL minimization problem is reformulated into the evidence lower bound (ELBO): \mathrm{ELBO}=\mathbb{E}_{q(w\mid\theta)}\!\left[\log p(D\mid w)\right]-\mathrm{KL}\!\left(q(w\mid\theta)\,\|\,p(w)\right), (2) as shown in Eq. ( 2 ). Optimizing the ELBO enables BNNs to learn weight distributions while providing uncertainty-aware predictions. II.2 Turbulence model correction The shear stress transport (SST) k-\omega model is widely used for turbulence modeling due to its ability to blend the strengths of the k-\varepsilon and k-\omega models. However, its reliance on Boussinesq’s assumption, which models the Reynolds stress tensor as proportional to the mean strain rate, introduces errors—particularly in separated flows where turbulence anisotropy plays a significant role. The Reynolds stress tensor \tau_{ij} can be decomposed into isotropic and anisotropic components: \tau_{ij}=2k\left(b_{ij}+\tfrac{1}{3}\,\delta_{ij}\right), (3) as shown in Eq. ( 3 ), where b_{ij}=-\nu_{t}/k\,S_{ij} is the normalized anisotropic tensor, \nu_{t} is the eddy viscosity, k is the turbulent kinetic energy (TKE), and \delta_{ij} is the Kronecker delta. To address these limitations, this study augments the baseline SST model by incorporating two correction terms: b_{ij}^{\Delta} , which modifies b_{ij} , and P_{k}^{\Delta} , which adjusts TKE production P_{k} The governing equations of the augmented SST model are expressed as: \frac{Dk}{Dt}=\nabla\cdot(\nu_{t}\nabla k)+P_{k}+P_{k}^{\Delta}-\beta^{*}\,\omega k,\qquad\frac{D\omega}{Dt}=\nabla\cdot(\nu_{t}\nabla\omega)+\gamma\,(P_{k}+P_{k}^{\Delta})/\nu_{t}-\beta\,\omega^{2}, (4) as shown in Eq. ( 4 ), where \beta and \beta^{*} are model parameters. For derivation and implementation details, refer to 6 . The correction terms b_{ij}^{\Delta} and P_{k}^{\Delta} are computed using the k -corrective frozen RANS approach 9 . In this approach, values of k , U , and b_{ij} are extracted from high-fidelity (HF) data, and the \omega -equation is solved iteratively while freezing other variables to compute deviations in the anisotropic stress tensor and TKE production. Following the Cayley–Hamilton theorem, b_{ij}^{\Delta} is formulated using a tensor basis expansion, providing a more flexible representation of turbulence anisotropy. Similarly, the correction term P_{k}^{\Delta} follows an analogous formulation, where the tensor basis functions (T_{ij}^{(n)}) and invariants (I_{m}) of the strain rate (S_{ij}) and rotation rate (\omega_{ij}) tensors define their structure. This correction term also includes an additional term in which \varepsilon=\omega k and \alpha_{4} is a function of the invariants. b_{ij}^{\Delta}=\sum_{n=1}^{N}T_{ij}^{(n)}\alpha_{n}(I_{m}),\qquad P_{k}^{\Delta}=2k\,b_{ij}^{\Delta}\frac{\partial U_{i}}{\partial x_{j}}+\alpha_{4}\varepsilon, (5) By incorporating b_{ij}^{\Delta} and P_{k}^{\Delta} , the model enhances turbulence representation in regimes where conventional closures lack accuracy as shown in equation (5). II.3 Zonal classification A key aspect of the proposed methodology is the identification of regions where turbulence model corrections are necessary. Instead of applying corrections globally, a classification method is introduced to selectively modify the turbulence closure in flow regions dominated by shear effects. To achieve this, we adopt the relative importance term analysis (RITA) classifier 4 , inspired by recent clustering approaches 3 ; 10 used to identify dominant physical processes in turbulent flows. The classification is based on the transport equation of TKE, where the relative contributions of production (P_{k}) , dissipation (D_{k}) , convection (C_{k}) , and diffusion (D_{f,k}) terms determine the turbulence dynamics at each point in the domain. RITA classifier evaluates the importance of these terms by computing relative ratios \mathrm{RITA}_{P_{k}/D_{k}}=\frac{|D_{k}|}{(|P_{k}|+|D_{k}|)},\qquad TI=\frac{k}{(k+0.5\,\|U\|^{2})}, (6) as demonstrated in equation (6). Cells satisfying the conditions \mathrm{RITA}_{P_{k}/D_{k}}<0.55 and TI\geq 0.12 are classified as part

Read the complete text