betaFeedback

WorksPhysics and Astronomy

Bayesian Kolmogorov-Arnold networks for trustworthy Reynolds-stress closures

Abstract

Data-driven turbulence closures can improve Reynolds-averaged Navier-Stokes (RANS) predictions, but their use is limited by two issues: learned constitutive relations are often difficult to inspect, and predictions typically lack reliable uncertainty estimates. We develop a Bayesian tensor-basis Kolmogorov-Arnold network (BKAN) for Reynolds-stress anisotropy that addresses both within a single closure. Learnable univariate spline functions on network edges provide an inspectable representation that can, where sufficiently simple, be approximated symbolically. Variational posteriors over spline coefficients yield a predictive distribution separating epistemic uncertainty, due to limited training data, from aleatoric uncertainty, the irreducible scatter of high-fidelity targets at given inputs. Galilean and rotational invariance, symmetry, and tracelessness are enforced through the tensor basis, while physical realisability is treated separately through soft regularisation. A controlled regression benchmark assesses predictive accuracy, calibration, and Bayesian parameterisation. Placing the posterior on spline coefficients gives the best overall balance of accuracy, calibration, parameter cost, and interpretability. Applied to Reynolds-stress anisotropy, the BKAN is accurate in distribution, with calibrated, informative uncertainty that tends to identify larger prediction errors. Under geometric extrapolation, however, the model becomes overconfident: epistemic uncertainty increases but does not fully account for growing prediction errors. Finally, the framework is extended to three-dimensional wind-farm wakes, where the learned correction transfers to unseen layouts and posterior samples are propagated through the RANS solver, with the aleatoric component injected as a spatially correlated field, to obtain uncertainty bounds on the predicted wake flow.

Introduction

Reynolds-averaged Navier–Stokes (RANS) modelling remains a practical basis for turbulent flow prediction when repeated high-fidelity simulations are computationally prohibitive. A central limitation, however, lies in the closure of the Reynolds stresses. Most commonly used models rely on the linear eddy viscosity assumption, which relates the Reynolds-stress anisotropy directly to the mean strain rate through the Boussinesq hypothesis. This representation is computationally efficient and numerically robust, but its constitutive form cannot reproduce important stress responses that arise in flows with separation, strong streamline curvature, or secondary motion. The resulting modelling errors are therefore often structural: adjusting closure coefficients may alter the prediction, but cannot recover physical behaviour that is absent from the model form itself [1].

Data-driven turbulence closures have been developed to address some of these structural limitations. Using high-fidelity data, such approaches can learn corrections to the Reynolds-stress anisotropy or to the production of turbulent kinetic energy that are not represented by conventional models. Tensor-basis neural networks [2] embed Galilean and rotational invariance into a learned anisotropy mapping. Field inversion and machine learning approaches [3] infer spatially varying model corrections, while sparse symbolic regression methods [4, 5] seek explicit algebraic closures from data. These approaches have improved both a priori and a posteriori predictions for a range of canonical turbulent flows. Broader perspectives on data-driven turbulence modelling and RANS model uncertainty are given by [6] and [7], respectively.

Despite this progress, two limitations remain important for the practical use of learned closures. The first is interpretability. Neural closures can be difficult to inspect, making it hard to determine which relationships have been learned, whether these relationships are physically meaningful, and how they behave outside the training regime. The second is predictive confidence. Deterministic closures provide point predictions without indicating how reliable those predictions are when the model is applied to conditions that differ from those represented in the training data.

When a learned closure does report uncertainty, it is useful to distinguish its two sources, because they carry different information for a turbulence model and behave differently at deployment, as sketched in Figure 1. Aleatoric uncertainty is the irreducible scatter of the high-fidelity target at a given set of input features: at fixed mean-flow invariants the reference anisotropy still varies, owing to noise in the data and to physical dependencies that the features do not capture. It is a property of the data and does not diminish as more data of the same kind are gathered. Epistemic uncertainty, by contrast, is uncertainty in the learned model itself, arising from finite and unevenly distributed training data; it is small where the data constrain the closure and grows where they do not, for example on an unseen geometry, and can in principle be reduced by broadening the training set. For a RANS closure the distinction is practical: epistemic uncertainty flags when a prediction is an extrapolation and should be trusted less, whereas aleatoric uncertainty sets a floor on how closely any closure of this form can match the reference. In this work we make the separation explicit and quantify both components, defining them formally in § 2.

Two kinds of predictive uncertainty for a data-driven closure

Figure 1. The two kinds of predictive uncertainty for a data-driven closure, sketched for a scalar correction g(λ)g(\lambda) against an input feature. (a) Aleatoric: the target scatters about the mean at fixed input; the predictive band follows this scatter and is not reduced by more data. (b) Epistemic: data constrain the model only over part of the input range, so the band is tight there and widens where data are absent.

These two issues have largely been studied separately. Bayesian and ensemble closures [8], together with physics-based eigenspace perturbation methods [9, 10], provide estimates of model uncertainty, but learned closures with uncertainty quantification often remain difficult to interpret. Moreover, their uncertainty is commonly assessed primarily through interval coverage. Coverage alone does not show whether larger predicted uncertainty corresponds to larger model error, or whether the uncertainty remains reliable under extrapolation. Conversely, symbolic methods and, more recently, Kolmogorov–Arnold turbulence closures [11, 12] provide more inspectable model forms, but are generally deterministic and therefore do not quantify predictive confidence. In particular, [12] identifies uncertainty quantification within the tensor-basis Kolmogorov–Arnold framework as an open problem.

An important precedent is the sparse Bayesian learning of explicit algebraic Reynolds-stress models [13]. In that approach, frame-invariant tensor polynomials provide an interpretable closure form, while a posterior distribution over the model coefficients provides uncertainty quantification that can also be propagated through the flow solver. The functional form, however, is selected from a prescribed dictionary of candidate tensor terms, and the Bayesian treatment is applied to the coefficients of a model that is linear in the selected features. A remaining question is whether the constitutive functions themselves can be learned while retaining an inspectable model structure and a predictive distribution. The resulting uncertainty should also be assessed using criteria that go beyond nominal interval coverage.

Kolmogorov–Arnold networks (KANs) [14] offer a useful model structure for this purpose. Unlike standard neural networks, which combine learned linear transformations with fixed nonlinear activation functions, KANs place learnable univariate functions on the network edges. These functions are typically represented by splines and can be inspected directly or, when their structure is sufficiently simple, approximated by symbolic expressions. This has motivated several interpretable and symbolic KAN formulations for scientific machine learning [15, 16]. The same property is attractive for tensor-basis turbulence closures, where the learning problem reduces to determining a relatively small set of scalar coefficient functions of invariant flow features.

Standard KANs, however, remain deterministic. Like conventional multilayer perceptrons (MLPs), they provide point predictions without quantifying predictive uncertainty. A Bayesian treatment addresses this limitation by assigning posterior distributions to selected model parameters. These posteriors induce distributions over the learned edge functions and, in turn, over the model predictions, allowing epistemic uncertainty associated with the learned closure to be distinguished from aleatoric uncertainty associated with the data. Bayesian variants of KANs have recently been introduced in the machine learning literature [17, 18], but their use as physically structured turbulence closures has not yet been examined. It also remains unclear which KAN parameters should carry the posterior and how the resulting uncertainty behaves when the model is applied outside the distribution represented by the training data.

The Bayesian Kolmogorov–Arnold network (BKAN) developed in this work extends the deterministic KAN in the same sense that a Bayesian neural network (BNN) extends a conventional MLP. BNNs have previously been used for turbulence closures with uncertainty quantification [8, 19], but their internal mappings remain difficult to interpret. Here, we combine the probabilistic treatment of Bayesian neural closures with the inspectable edge functions of KANs and the invariant structure of a tensor-basis representation.

Specifically, we formulate a Bayesian tensor-basis KAN for Reynolds-stress anisotropy that combines interpretable spline functions with a posterior predictive distribution, while Galilean and rotational invariance, symmetry, and tracelessness are enforced through the tensor-basis representation. We examine whether the posterior should be placed on the spline coefficients, the edge amplitudes, or both, and compare these choices in terms of calibration, parameter count, and interpretability. The quality of the predicted uncertainty is assessed not only through calibration, but also through sharpness and sparsification-based error ranking. This provides a more direct test of whether regions assigned high uncertainty are also regions where the model makes larger errors, and how this relationship changes under extrapolation. The closure is then evaluated across a hierarchy of turbulent flows, from canonical planar and duct configurations to three-dimensional wind farm wakes. For the wind farm cases, posterior samples are also propagated through the flow solver to quantify uncertainty in predicted flow quantities.

The remainder of the paper is organised as follows. § 2 presents the BKAN closure, its tensor-basis formulation, and the decomposition of predictive uncertainty, with the full variational formulation given in Appendix A.§ 1. § 3 evaluates the underlying Bayesian regressor on a controlled benchmark. § 4 considers Reynolds-stress anisotropy prediction in both in-distribution and cross-geometry settings. § 5 extends the formulation to three-dimensional wind farm wakes and solver-propagated uncertainty. § 6 discusses the implications, the trust boundary, and the limitations of the approach. § 7 summarises the main conclusions.

A Bayesian Kolmogorov–Arnold network for tensor-basis closures

The closure problem and tensor-basis representation

The Reynolds-stress anisotropy tensor is defined as

bij=uiuj‾2k−13δij,k=12uiui‾,(1)b_{ij} = \frac{\overline{u_i u_j}}{2k} - \frac{1}{3}\delta_{ij}, \qquad k = \frac{1}{2}\overline{u_i u_i}, \tag*{(1)}

Here uiuj‾\overline{u_i u_j} is the Reynolds-stress tensor, uiu_i the fluctuating velocity, kk the turbulent kinetic energy, and δij\delta_{ij} the Kronecker delta; bijb_{ij} is the normalised, trace-free anisotropy tensor obtained by scaling the Reynolds stress by 2k2k and removing its isotropic part. It is the closure quantity considered in this work. Linear eddy viscosity models often represent this anisotropy inadequately because they restrict its relation to the mean strain rate. Following [20], and assuming that the anisotropy is a local function of the normalised mean strain- and rotation-rate tensors, S^\widehat{\mathbf{S}} and W^\widehat{\mathbf{W}}, alone, a general frame-invariant relation between bijb_{ij} and these tensors can be expressed as a finite tensor polynomial expansion,

bij(x)=∑n=1Ngn(λ1,…,λm)Tij(n)(S^,W^),(2)b_{ij}(\mathbf{x}) = \sum_{n=1}^{N} g_n(\lambda_1,\ldots,\lambda_m) T_{ij}^{(n)}(\widehat{\mathbf{S}},\widehat{\mathbf{W}}), \tag*{(2)}

where Tij(n)T_{ij}^{(n)} are integrity-basis tensors and λp\lambda_p are scalar invariants formed from S^\widehat{\mathbf{S}} and W^\widehat{\mathbf{W}}. For two-dimensional mean flows, three basis tensors and two invariants are sufficient for a perfectly general representation. In three dimensions, the full representation contains ten basis tensors and five invariants. The tensors and invariants used in this work are given in Appendix A.§ 2.

Equation (2) separates the tensorial structure of the closure from the unknown scalar coefficient functions gng_n. Since the basis tensors are constructed from S^\widehat{\mathbf{S}} and W^\widehat{\mathbf{W}} and the coefficients depend on invariant scalar quantities, the resulting representation preserves the required frame invariance while enforcing symmetry and tracelessness by construction. The learning problem is therefore reduced to determining a relatively small set of scalar coefficient functions rather than learning the individual components of bijb_{ij} independently.

For the learned closure, we supplement the tensor invariants with three additional Galilean-invariant scalar markers: a measure of the balance between strain and rotation, a turbulence-intensity measure, and a streamline pressure-gradient marker. These quantities provide information that is not contained in the pointwise tensor invariants alone. In what follows, λ∈Rm\lambda\in\mathbb{R}^{m} denotes the complete scalar feature vector of dimension mm supplied to the network, consisting of the tensor invariants together with these additional markers. Their definitions are given in Appendix A.§ 2.

The Kolmogorov–Arnold architecture

A KAN represents a multivariate function through compositions of learnable univariate functions placed on the network edges [14]. This differs from an MLP, where learned linear transformations are combined with fixed nonlinear activation functions. In a KAN, each edge carries a function of the form

ϕ(x)=wbb(x)+wsϕs(x),ϕs(x)=∑cθcBc(x),(3)\phi(x) = w_b b(x) + w_s \phi_s(x), \qquad\phi_s(x) = \sum_c \theta_c B_c(x), \tag*{(3)}

where b(x)b(x) is a fixed basis function, Bc(x)B_c(x) denotes the ccth B-spline basis function on a prescribed knot grid, and θc\theta_c is its trainable coefficient. The amplitudes wbw_b and wsw_s scale the fixed and spline contributions, respectively. Because each edge function is one-dimensional, it can be inspected directly over its input range. When a learned function has a sufficiently simple shape, it can also be approximated by a closed-form expression. We use this property for the symbolic distillation in § 3, following the symbolic regression procedure of [21].

In the present closure, the feature vector λ\lambda is mapped by the BKAN to the scalar tensor-basis coefficients gng_n. These coefficients are combined with the fixed Pope basis tensors through (2) to reconstruct the Reynolds-stress anisotropy. A second output head uses the same input features to predict the input-dependent variance required by the heteroscedastic likelihood. The resulting architecture, with a shared BKAN representation and separate coefficient and variance heads, is illustrated in Figure 2.

Schematic of the Bayesian tensor-basis KAN closure

Figure 2. Schematic of the Bayesian tensor-basis KAN closure. The invariant feature vector λ\boldsymbol{\lambda} is mapped by a shared KAN with learnable spline edges to two output heads: the tensor-basis coefficients gng_n and the heteroscedastic variance. The coefficients are combined with the fixed Pope basis tensors to reconstruct bij=∑n=1NgnTij(n)b_{ij} = \sum_{n=1}^{N} g_n T_{ij}^{(n)}, posterior sampling provides the epistemic uncertainty and the variance head the aleatoric uncertainty.

Variational Bayesian treatment

To quantify the epistemic uncertainty in the learned closure—the uncertainty in the model parameters that arises from training on a finite data set—we place a mean-field Gaussian variational posterior qψq_\psi over the subset of model parameters treated as Bayesian variables, denoted θ∈RP\theta\in\mathbb{R}^{P} with PP the number of such parameters. The posterior is parameterised by the variational parameters ψ={(μi,σi)}i=1P\psi= \{(\mu_i,\sigma_i)\}_{i=1}^{P}, the mean and standard deviation of each Gaussian factor, which are optimised during training. The model is trained by maximising the evidence lower bound (ELBO),

L(ψ)=Eqψ[log⁡p(D∣θ)]−DKL(qψ(θ)∥p(θ)),(4)\mathcal{L}(\psi) = \mathbb{E}_{q_\psi}\left[\log p(D \mid\theta)\right] - D_{\mathrm{KL}}\left(q_\psi(\theta) \parallel p(\theta)\right), \tag*{(4)}

following the Bayes-by-backprop formulation of [22]. Here, DKLD_{\mathrm{KL}} is the Kullback–Leibler divergence between the variational posterior and the prior. The local reparameterisation trick [23] is used to reduce gradient variance through the spline activations. For the spline branch of an edge with Gaussian coefficients θk∼N(μk,σk2)\theta_k \sim\mathcal{N}(\mu_k,\sigma_k^2), and because the spline ϕs(x)\phi_s(x) is linear in those coefficients, its output at a given input is itself Gaussian, with mean ∑kBk(x)μk\sum_k B_k(x)\mu_k and variance ∑kBk(x)2σk2\sum_k B_k(x)^2\sigma_k^2. The spline output can therefore be sampled directly from this scalar distribution rather than by sampling every coefficient individually. The full derivation is given in Appendix A.§ 1.

The data term is based on a heteroscedastic Gaussian likelihood over the dd unique components of the anisotropy tensor. Its negative log-likelihood (NLL) is

NLL=12∑c=1d[log⁡σc2(λ)+(bc−b^c)2σc2(λ)],(5)\mathrm{NLL} = \frac{1}{2}\sum_{c=1}^{d}\left[\log\sigma_c^2(\boldsymbol{\lambda}) + \frac{(b_c-\hat{b}_c)^2}{\sigma_c^2(\boldsymbol{\lambda})}\right], \tag*{(5)}

where bcb_c and b^c\hat{b}_c denote the reference and predicted values of the ccth unique anisotropy component, respectively, and σc2(λ)\sigma_c^2(\boldsymbol{\lambda}) is its predicted aleatoric variance. Allowing this variance to depend on the input gives the likelihood a spatially varying width rather than assuming a constant noise level.

In practice, we minimise the negative ELBO together with the realisability regularisation introduced below,

J(ψ)=NLL⁡(bij∣λ)+αR(b^ij)+βDKL(qψ∥p)N,(6)J(\psi) = \operatorname{NLL}(b_{ij} \mid\lambda) + \alpha R(\hat{b}_{ij}) + \beta\frac{D_{\mathrm{KL}}(q_{\psi}\mathbin{\|}p)}{N}, \tag*{(6)}

where NN is the number of training points, α\alpha controls the weight of the realisability penalty, and β\beta controls the contribution of the KL term. During training, β\beta is gradually increased from zero, allowing the network to first fit the data before the posterior is progressively regularised toward the prior.

The variance of the Gaussian likelihood is predicted by the second output head of the BKAN using the same feature vector λ\lambda. The head predicts an input-dependent heteroscedastic log variance, which is mapped to the positive variance σc2(λ)\sigma_c^2(\lambda) used in (5). This allows the aleatoric uncertainty to vary with the local flow state [24, 25]. In the present setting this aleatoric component represents the irreducible scatter of the high-fidelity target that remains once the invariant features are fixed, arising from noise in the reference data and from physical dependencies not captured by the chosen feature set. Unlike the epistemic term, it is not reduced by adding more training data of the same kind, and the heteroscedastic form lets it grow in regions, such as strong shear or separation, where the feature-to-anisotropy relationship is intrinsically more scattered. This variance is pointwise: the aleatoric uncertainty at a cell depends only on that cell’s features, and no correlation between neighbouring cells is modelled at this stage. Such spatial correlation becomes relevant only when the aleatoric component is propagated through the flow solver, where it is introduced explicitly (§ 5.3).

We consider three prior formulations for the Bayesian parameters: a Gaussian prior with an analytic KL term, the two-component scale-mixture prior of [22], and a Gamma hyperprior that provides automatic relevance determination [26]. Their effect on the results is small, and the Gaussian prior is therefore used unless stated otherwise. The corresponding KL expressions and the local reparameterisation derivation are given in Appendix A.§ 1.

Training begins with a short deterministic stage in which the posterior means are optimised before variational inference. We found this initialisation useful for resolving sharp spatial variations in the target fields. Further training details are given in Appendix A.§ 3, and the complete training and prediction procedure is summarised in Algorithm 1.

Choice of Bayesian parameters

The KAN edge in (3) contains two groups of parameters that can be treated probabilistically. The spline coefficients θc\theta_c determine the shape of the learned univariate functions, whereas the amplitudes wbw_b and wsw_s scale the fixed and spline contributions. We therefore consider three Bayesian parameterisations: a posterior over the spline coefficients only (b-coef), a posterior over the edge amplitudes only (b-scale), and a posterior over both sets of parameters (b-full).

The three variants differ in parameter count and in how uncertainty is represented within each edge function. Each Bayesian parameter is represented by a variational mean and log-standard-deviation. In b-coef, the spline coefficients {θc}\{\theta_c\} are treated as Bayesian variables while wbw_b and wsw_s remain deterministic. In b-scale, the amplitudes are Bayesian and the spline coefficients remain deterministic, while b-full places a posterior over both. Consequently, b-coef introduces two variational parameters for each spline coefficient, b-scale introduces two for each Bayesian edge amplitude, and b-full includes both contributions.

This distinction also affects interpretability. Since the spline coefficients determine the shape of each learned univariate function, placing a posterior over them produces a distribution over

Algorithm 1 Training and prediction of the Bayesian tensor-basis KAN closure.

  1. Inputs: invariant features {λ(p)}\{\lambda^{(p)}\}, basis tensors {T(n,p)}\{T^{(n,p)}\}, labels {bij(p)}\{b_{ij}^{(p)}\}; pre-training epochs E0E_0; variational epochs E1E_1; anneal length EaE_a

  2. initialise variational means μ\boldsymbol{\mu} and log-standard-deviations ρ\boldsymbol{\rho}

  3. for e=1,…,E0e=1,\ldots,E_0 do ▷ deterministic pre-training of the means

  4. gn←KAN⁡(λ;μ)g_n \leftarrow\operatorname{KAN}(\lambda;\boldsymbol{\mu}); b^ij←∑ngnTij(n)\hat{b}_{ij}\leftarrow\sum_n g_nT_{ij}^{(n)}

  5. update μ\boldsymbol{\mu} by gradient descent on ∥b^ij−bij∥2\lVert\hat{b}_{ij}-b_{ij}\rVert^2

  6. end for

  7. for e=1,…,E1e=1,\ldots,E_1 do ▷ variational inference

  8. β←min⁡(1,e/Ea)\beta\leftarrow\min(1,e/E_a) ▷ KL annealing

  9. for each mini-batch do

  10. sample spline activations by local reparameterisation, (A.4)

  11. b^ij←∑ngnTij(n)\hat{b}_{ij}\leftarrow\sum_n g_nT_{ij}^{(n)}; log⁡σij2←variance-head⁡(λ)\log\sigma_{ij}^2\leftarrow\operatorname{variance\text{-}head}(\lambda)

  12. J←NLL⁡+αR(b^ij)+βDKL/NJ\leftarrow\operatorname{NLL}+\alpha R(\hat{b}_{ij})+\beta D_{\mathrm{KL}}/N ▷ (6)

  13. update (μ,ρ)(\boldsymbol{\mu},\boldsymbol{\rho}) and the variance head using Adam on JJ

  14. end for

  15. end for

  16. Predict: draw SS posterior samples {g(s)}\{g^{(s)}\}, evaluate bˉij\bar{b}_{ij} at the posterior-mean coefficients, and compute the epistemic variance Var⁡s(bij(s))\operatorname{Var}_s(b_{ij}^{(s)}) and aleatoric variance Es[σij2]\mathbb{E}_s[\sigma_{ij}^2] via (7)

the function itself rather than only over its overall amplitude. Which parameterisation gives the most useful balance between predictive performance, uncertainty quality, parameter cost, and interpretability is not clear a priori. We therefore compare the three variants empirically in § 3.

Predictive distribution and its decomposition

At prediction time, the posterior over the coefficient functions induces a predictive distribution over bijb_{ij}. Drawing SS posterior samples θ(s)∼qψ\theta^{(s)}\sim q_\psi gives a corresponding set of reconstructed anisotropy tensors bij(s)b_{ij}^{(s)} through (2). Their spread represents epistemic uncertainty, while the input-dependent variance predicted by the variance head provides the aleatoric component. By the law of total variance, the predictive variance is decomposed as

Var⁡[bij]⏟predictive=Var⁡s[bij(s)]⏟epistemic+Es[σij2(λ)]⏟aleatoric.(7)\underbrace{\operatorname{Var}[b_{ij}]}_{\text{predictive}}=\underbrace{\operatorname{Var}_s[b_{ij}^{(s)}]}_{\text{epistemic}}+\underbrace{\mathbb{E}_s[\sigma_{ij}^2(\lambda)]}_{\text{aleatoric}}. \tag*{(7)}

The epistemic term reflects uncertainty in the learned closure and is expected to increase when predictions rely on regions of the input space that are poorly constrained by the training data. The aleatoric term represents the input-dependent scatter captured by the heteroscedastic likelihood. This decomposition therefore separates uncertainty associated with the learned model from that represented by the likelihood.

For the reported point field, we evaluate the model at the posterior-mean coefficients rather than averaging the SS sampled predictions. This avoids Monte-Carlo noise in the displayed prediction, while the epistemic variance is estimated from the posterior samples. The corresponding moments are accumulated online over S=50S=50–100100 draws, avoiding storage of an array of size O(S×N)\mathcal{O}(S \times N). The quality of the resulting uncertainty estimate, including its calibration, sharpness, and ability to identify regions of larger prediction error, is examined in § 4.

Physical realisability

The anisotropy tensor is physically realisable when its eigenvalues remain within the interval [−13,23]\left[-\frac{1}{3}, \frac{2}{3}\right], or equivalently when the corresponding barycentric coordinates are non-negative [27]. To discourage predictions outside this admissible region, we introduce the differentiable penalty

R(bij)=⟨max⁡(0,λmax⁡−23)2+max⁡(0,−13−λmin⁡)2⟩(8)R(b_{ij}) = \left\langle\max\left(0,\lambda_{\max}-\frac{2}{3}\right)^2+\max\left(0,-\frac{1}{3}-\lambda_{\min}\right)^2\right\rangle \tag*{(8)}

where the average is taken over the training points and λmin⁡\lambda_{\min} and λmax⁡\lambda_{\max} denote the smallest and largest eigenvalues of the predicted anisotropy tensor. The penalty follows the realisability treatment of [11].

Geometrically, the ordered eigenvalues determine a point in the barycentric representation, with realisable anisotropy states lying inside the admissible triangle. Equation (8) penalises predictions whose extreme eigenvalues leave the admissible interval, with the strength of the regularisation controlled by α\alpha. Importantly, the tensor-basis representation enforces invariance, symmetry, and tracelessness, but does not by itself guarantee physical realisability. The penalty is therefore used as a soft regulariser rather than a hard constraint. Its empirical effect is examined in § 4 and discussed further in § 6. The penalty is applied only to the bounded anisotropy target and not to the signed correction considered in § 5.

Reference models

To isolate the effect of the Kolmogorov–Arnold architecture, we compare the BKAN with a matched tensor-basis BNN baseline. The BNN uses a mean-field variational MLP to map the same feature vector λ\lambda to the tensor-basis coefficients gng_n, followed by the same fixed tensor-basis reconstruction in (2). It also uses the same heteroscedastic variance head, realisability penalty, and training protocol. The two models therefore differ primarily in the function approximator used to predict the coefficients gng_n: a KAN in the proposed model and an MLP in the BNN baseline. The BNN baseline follows the Bayesian tensor-basis closure framework of [8]. Related Bayesian approaches to turbulence-model correction have been developed by [13] and, concurrently with the present work, by [19]. For the controlled regression study in § 3, we additionally compare the BKAN with a deterministic MLP, a deterministic KAN, Monte Carlo dropout [28], and deep ensembles [29], using matched parameter budgets.

Evidence from controlled regression

Before applying the BKAN to turbulence modelling, we first assess its behaviour as a Bayesian regressor on a standard benchmark. The purpose is to determine whether the underlying model can achieve competitive predictive accuracy at a comparable parameter budget, provide well-calibrated uncertainty estimates, and retain the interpretability associated with the KAN structure. We use nine datasets from the regression benchmark of [30], which is widely used for evaluating probabilistic neural-network models.

The complete experimental protocol, dataset descriptions, and per-dataset results are given in Appendix B.§ 1. The comparison is performed at matched parameter budget: for each case, the KAN width is selected such that its number of trainable parameters does not exceed that of the reference BNN. This reduces the extent to which performance differences can be attributed simply to model size. We report three metrics. The root-mean-square error measures predictive accuracy,

RMSE=1N∑i=1N(yi−y^i)2(9)\mathrm{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(y_i-\hat{y}_i)^2} \tag*{(9)}

with y^i\hat{y}_i the predictive mean. The test log-likelihood measures the quality of the full predictive distribution,

LL=1N∑i=1Nlog⁡N(yi∣y^i,σi2),(10)\mathrm{LL} = \frac{1}{N}\sum_{i=1}^{N}\log\mathcal{N}(y_i \mid\hat{y}_i,\sigma_i^2), \tag*{(10)}

rewarding predictions that are accurate and appropriately uncertain, with σi2\sigma_i^2 the predictive variance. Calibration is measured by the expected calibration error: for nominal central-interval levels {pk}k=1K\{p_k\}_{k=1}^{K} with empirical coverages p^k\hat{p}_k (the fraction of test targets inside the corresponding predictive interval),

ECE=1K∑k=1K∣p^k−pk∣,(11)\mathrm{ECE} = \frac{1}{K}\sum_{k=1}^{K}\lvert\hat{p}_k-p_k\rvert, \tag*{(11)}

following [31], so that ECE = 0 corresponds to intervals with their nominal coverage. The BKAN is compared with a variational BNN, a deterministic MLP, a deterministic KAN, Monte Carlo dropout, and deep ensembles.

Accuracy and calibration

At matched parameter budget, the BKAN achieves predictive accuracy and test log-likelihood comparable to those of the BNN across the benchmark suite, while also providing well-calibrated uncertainty estimates, shown in Figure 3. The comparison does not indicate uniform superiority: on several datasets the two Bayesian models perform similarly, while on others the BNN gives lower predictive error. The full per-dataset results are reported in Appendix B.§ 1.

Figure 3. Calibration error (ECE) on the regression benchmark at matched parameter budget, averaged over the nine datasets. Predictive error and test log-likelihood are near-identical across models and are omitted here (full per-dataset results in Appendix B.§ 1). The BKAN is among the best calibrated, comparable to deep ensembles.

Averaged over the nine datasets, the BKAN is among the two best-calibrated of the six models considered, with a mean calibration error of approximately 0.07. This is comparable to the deep-ensemble result and lower than the BNN value of approximately 0.09, while using a matched or smaller parameter budget. Predictive error and test log-likelihood are also comparable to the BNN on most datasets. For the purpose of the present study, these results show that the BKAN provides competitive regression accuracy together with calibrated predictive uncertainty. Its behaviour on the turbulence cases can therefore be examined without the benchmark performance itself being a limiting factor.

Bayesian parameterisation study

The benchmark also addresses the design question introduced in § 2.4: which KAN parameters should carry the posterior. Figure 4 compares the three parameterisations across five benchmark datasets. Placing the posterior only on the edge amplitudes (b-scale) gives the least reliable calibration, with calibration errors reaching approximately 0.4 on some datasets, without providing a consistent improvement in predictive accuracy. This suggests that uncertainty in the per-edge amplitudes alone is not sufficient to represent the predictive uncertainty of the learned functions.

Figure 4. Effect of the Bayesian parameterisation across five benchmark datasets. The edge-amplitude posterior (b-scale) gives the least reliable calibration, while the spline-coefficient posterior (b-coef) provides the most favourable balance between calibration, predictive accuracy, parameter cost, and interpretability. The full posterior (b-full) generally produces the largest epistemic spread, but at a higher parameter cost.

Placing the posterior on the spline coefficients (b-coef) gives the most favourable overall balance. It remains well calibrated across the datasets and places the posterior directly on the parameters that determine the shape of each learned edge function. The full parameterisation (b-full), in which both the spline coefficients and edge amplitudes are Bayesian, generally produces the largest epistemic spread. This additional flexibility comes at a higher parameter cost, without a consistent improvement in calibration over b-coef.

Based on these results, we use b-coef as the default Bayesian parameterisation throughout the remainder of the paper. The turbulence-anisotropy ablation in Appendix § 12 (Table Appendix C2) supports the same choice: b-coef gives the lowest RMSE and calibration error, whereas both alternative parameterisations produce larger epistemic uncertainty and less favourable overall predictive performance.

Interpretability with uncertainty

We next examine the interpretability of the Bayesian KAN representation using the tensor-basis closure for the periodic-hill flow. Each edge function in (3) is a one-dimensional curve that can be inspected directly. Under the b-coef parameterisation, the posterior over the spline coefficients also induces a distribution over the learned edge function. Its mean can therefore be examined together with a posterior uncertainty band and, when the learned shape is sufficiently simple, approximated by a closed-form expression.

Unlike symbolic KANs that incorporate symbolic structure directly into training through gated primitive selection [16], the symbolic distillation used here is performed post hoc on the learned functions. The recovered expression can therefore be compared directly with the posterior mean and its uncertainty band.

For the minimal two-invariant periodic-hill model used for this analysis, each candidate expression is fitted and scored only over the range in which the training data support the learned curve, namely the 2–98% band of the standardised invariant shaded in Figure 5. This restriction is important. The swept range extends beyond the data, and approximately 87% of the variance of a coefficient curve lies in the extrapolated tails, so a candidate that merely follows those tails can reach R2≈0.98R^2 \approx0.98 while explaining almost nothing where the data constrain the function.

Symbolic recovery of the tensor-basis coefficient functions

Figure 5. Symbolic recovery of the tensor-basis coefficient functions gn(λ1)g_n(\lambda_1) for the minimal two-invariant periodic-hill closure, sweeping the dominant invariant λ1\lambda_1 with λ2\lambda_2 held at its median. Each panel shows the posterior mean (solid blue), the ±2σ\pm2\sigma posterior band (shaded), and the best closed-form fit (crimson dashed) over the 2–98% data range (grey), within which the candidates are fitted and scored. Only the third coefficient admits a simple form; the fits and R2R^2 values are discussed in § 3.

Within the supported range, the third coefficient function is well represented by

g3(λ1)≈0.24sin⁡(0.90λ1+1.59)−0.20,(12)g_3(\lambda_1) \approx0.24\sin(0.90\lambda_1 + 1.59) - 0.20, \tag*{(12)}

which reproduces the posterior-mean function with R2=0.96R^2 = 0.96. Applying the same functional form to the upper and lower ±2σ\pm2\sigma curves gives R2=0.95R^2 = 0.95 in both cases, so the recovered expression remains consistent across the posterior band rather than only at its mean. The two remaining coefficients do not admit a comparably simple closed form over the same range: for both, the best candidate in the library is linear, with R2=0.06R^2 = 0.06 for g1g_1 and R2=0.14R^2 = 0.14 for g2g_2.

Because the tensor basis is partly collinear in two dimensions, the allocation of the predicted anisotropy among the individual coefficient functions is not unique, and the distillation therefore depends on the initialisation. Repeating the analysis over thirteen random seeds gives thirty-nine coefficient fits, of which eight reach R2>0.9R^2 > 0.9 while twenty-eight fall below R2=0.3R^2 = 0.3. The leading coefficient never admits a compact form, its largest R2R^2 over all seeds being 0.33, whereas the third distils successfully for six of the thirteen seeds. Among the successful fits the recovered form is itself divided between sinusoidal and quadratic candidates. Figure 5 therefore shows a representative case (random seed 17) and should be read as an inspection of one learned coefficient representation rather than as a unique symbolic form of the complete closure. The result demonstrates that the internal scalar functions of the KAN can be inspected together with their posterior uncertainty and, where their structure is sufficiently simple, represented by compact analytical expressions, while also delimiting how often that simplicity is available.

Reynolds-stress anisotropy with quantified uncertainty

We now apply the proposed closure to its intended task: a priori prediction of the Reynolds-stress anisotropy from RANS mean-flow inputs, using high-fidelity data as reference. We use the curated dataset of [32], which provides matched RANS and direct numerical simulation (DNS) or large eddy simulation (LES) fields for a family of turbulent flows; the cases and their train/test splits are detailed in Appendix B.§ 2. The planar cases considered here are periodic hills, a converging–diverging channel, and a curved backward-facing step, which involve separation and streamline curvature. The dataset also includes a square duct, where secondary flow of Prandtl’s second kind is driven by Reynolds-stress anisotropy and cannot be reproduced by a linear eddy-viscosity model.

For the planar flows, the three-term tensor basis is exact, while the square duct is represented using the full ten-term basis. Throughout this section, we distinguish between two prediction regimes. By a flow family we mean a group of flows that share the same geometry type and governing physics and differ only through a parameter such as the Reynolds number or a geometric ratio. In distribution then refers to interpolation within such a family, where the training and test data are held-out cells, Reynolds numbers, or parameter values drawn from the same family. Out of distribution refers to prediction on a geometry that belongs to no family represented in training and therefore provides a more demanding test of generalisation. The architecture, input features, and training procedure are summarised in Appendix A.§ 3. All in-distribution metrics are reported as means over three random seeds.

Accuracy in distribution

For the periodic-hill flow, the closure predicts the Reynolds-stress anisotropy with an RMSE of 0.046 ±\pm 0.001 over three random seeds. The small variation between seeds indicates limited sensitivity to the model initialisation. The predicted component fields reproduce the main spatial structures of the DNS reference through the separated shear layer and recirculation region, with the largest discrepancies remaining locally around the more strongly varying parts of the flow, shown in Figure 6.

In-distribution anisotropy prediction for the periodic-hill flow

Figure 6. In-distribution anisotropy prediction for the periodic-hill flow: DNS reference and BKAN prediction for selected components. The predicted fields reproduce the main structures through the separated and recirculating regions.

The square duct provides a different test because its secondary flow is driven by Reynolds-stress anisotropy. In particular, gradients of the cross-plane normal-stress difference together with the shear component b23b_{23} contribute to the generation of streamwise vorticity. These effects cannot be represented by a linear eddy-viscosity model, which constrains the anisotropy to be proportional to the mean strain rate. The BKAN reproduces the spatial structure and sign of the shear component b23b_{23} that drives the secondary motion, shown in Figure 7.

In-distribution prediction for the square duct, showing the shear component $b_{23}$

Figure 7. In-distribution prediction for the square duct (ten-term basis), showing the shear component b23b_{23} that drives the secondary motion: DNS reference, BKAN prediction, and pointwise absolute error.

To examine the effect of the Kolmogorov–Arnold architecture, we compare the BKAN with the tensor-basis BNN introduced in § 2.7. The two models use the same tensor-basis reconstruction, realisability penalty, heteroscedastic variance head, and training protocol, and differ primarily in the function approximator used to map the input features to the coefficients gng_n. In distribution, the BKAN and the larger BNN give comparable predictive errors, with RMSE values close to 0.046, as shown in Figure 8. Their calibration differs more clearly: the BKAN has an ECE of approximately 0.06, compared with approximately 0.09 for the smaller BNN and 0.13 for the wider BNN. Increasing the BNN capacity therefore does not improve its predictive accuracy in this comparison and is accompanied by poorer calibration. The BKAN achieves comparable accuracy while retaining the inspectable KAN representation and giving the lowest in-distribution calibration error among the three models shown.

Figure 8. Comparison of the BKAN with two tensor-basis BNN baselines. The numbers in parentheses indicate the number of model parameters. All models use the same tensor-basis reconstruction, realisability penalty, heteroscedastic variance formulation, and training protocol. In distribution, the models give comparable predictive accuracy, while the BKAN has the lowest calibration error. The out-of-distribution results are discussed in § 4.3.

Calibration and informativeness of predictive uncertainty

Predictive accuracy alone does not establish whether the associated uncertainty estimate is useful. We therefore assess three complementary properties: calibration, sharpness, and informativeness.

Calibration measures whether the predicted intervals have their stated coverage. A reliability diagram plots the empirical coverage p^k\hat{p}_k against the nominal level pkp_k, and the expected calibration error of Eq. (11) summarises the departure from the diagonal (ECE = 0 for intervals with exactly their nominal coverage).

Sharpness measures how concentrated the predictive distribution is, independently of the reference values. We report the mean total predictive standard deviation,

Sharp⁡=1N∑i=1Nσtot,i,(13)\operatorname{Sharp} = \frac{1}{N}\sum_{i=1}^{N}\sigma_{\mathrm{tot},i}, \tag*{(13)}

with σtot,i\sigma_{\mathrm{tot},i} the total predictive standard deviation at point ii; smaller values indicate a more concentrated distribution. It is reported together with calibration because nominal coverage can otherwise be obtained simply by widening the intervals.

Informativeness measures whether the predicted uncertainty ranks the actual errors. Predictions are sorted by predictive uncertainty and progressively removed, giving a sparsification curve eunc(f)e_{\mathrm{unc}}(f) of mean remaining error against the fraction ff removed. This is compared with an oracle curve eoracle(f)e_{\mathrm{oracle}}(f) obtained by instead sorting on the true error; the area between them,

AUSE=∫01[eunc(f)−eoracle(f)] df,(14)\mathrm{AUSE} = \int_{0}^{1} [e_{\mathrm{unc}}(f) - e_{\mathrm{oracle}}(f)]\,\mathrm{d}f, \tag*{(14)}

is the area under the sparsification error. A random ordering provides an upper reference, and AUSE→0\to0 when the uncertainty ranks errors as well as the true error itself.

In distribution, the reliability curve lies close to the ideal diagonal and is slightly conservative, with empirical coverage generally exceeding the nominal coverage (Figure 9, top-left). The corresponding calibration error is ECE = 0.072, with a sharpness of 0.042. The sparsification result is also favourable: the BKAN curve remains substantially closer to the oracle than to the random ordering, with AUSE = 0.115 (Figure 9, bottom-left). This indicates that regions assigned larger uncertainty also tend to contain larger prediction errors.

Figure 9. Quality of the predictive uncertainty in distribution (left) and out of distribution (right). The top row shows reliability diagrams together with the corresponding sharpness, while the bottom row shows sparsification curves relative to oracle and random error rankings. In distribution, the BKAN is well calibrated and its uncertainty provides an informative ranking of prediction error. Out of distribution, calibration and error ranking deteriorate substantially; this behaviour is examined further in § 4.3.

The spatial structure of the uncertainty provides a complementary view. For the periodic-hill flow, the epistemic standard deviation of b12b_{12} is largest along the separated shear layer and around reattachment in Figure 10. These are also regions where the learned mapping is more strongly challenged by the local flow behaviour. The aleatoric and epistemic contributions have distinct spatial distributions, showing that the uncertainty estimate contains information beyond aggregate coverage alone.

Epistemic and aleatoric predictive standard deviation of the $b_{12}$ component for the periodic-hill flow

Figure 10. Epistemic (left) and aleatoric (right) predictive standard deviation of the b12b_{12} component for the periodic-hill flow. The epistemic component is concentrated along the separated shear layer and near reattachment, whereas the aleatoric contribution has a different spatial distribution.

The limits of extrapolation

We next consider prediction on a geometry that is absent from the training data. When the model is trained on a subset of the available flows and evaluated on a withheld geometry, the predictive accuracy deteriorates substantially. The cross-geometry RMSE increases to 0.116±\pm0.009,

compared with 0.046 ± 0.001 in distribution. The larger variation between random seeds also indicates greater sensitivity to the learned model under extrapolation. For the curved backward-facing step shown in Figure 11, the shear anisotropy generalises more successfully than the normal-stress components, although all components show larger errors than in the in-distribution case.

DNS and BKAN predictions for the unnormalised anisotropy components

Figure 11. Out-of-distribution prediction for the withheld curved backward-facing-step geometry, shown as the unnormalised anisotropy aij=2kbija_{ij}=2kb_{ij}: because the turbulent kinetic energy kk is small here, the normalised bijb_{ij} is dominated by noise. DNS reference and BKAN prediction for a11a_{11}, a12a_{12} and a22a_{22}; the shear component generalises more successfully than the normal components.

The uncertainty estimate also deteriorates out of distribution. The reliability curve falls well below the ideal diagonal, with ECE=0.319\mathrm{ECE} = 0.319, showing substantial under-coverage of the predictive intervals (Figure 9, top-right). At the same time, the sharpness changes only slightly, from 0.042 in distribution to 0.040 out of distribution. The loss of calibration therefore occurs without a corresponding increase in the width of the predictive distribution. The sparsification result also deteriorates markedly, with AUSE=0.665\mathrm{AUSE} = 0.665, compared with 0.115 in distribution (Figure 9, bottom-right). The predicted uncertainty consequently provides a much poorer ranking of the actual errors under geometric extrapolation.

The spatial epistemic uncertainty nevertheless responds to the change in geometry. It increases and is concentrated in regions where the withheld flow differs most strongly from those represented in training, shown in Figure 12. The aleatoric component, by contrast, follows the local magnitude of the correction and is more broadly distributed than the epistemic term, so the two carry complementary information: the epistemic map flags where the inputs are unfamiliar, whereas the aleatoric map reflects the intrinsic scatter of the target. The model therefore shows some sensitivity to extrapolation, but the magnitude of its epistemic uncertainty is insufficient to account for the resulting prediction error. This combination of increased epistemic uncertainty and substantial under-coverage defines a practical trust boundary for the present model: within the represented flow family, the uncertainty is calibrated and informative, whereas for unseen geometries it should be interpreted with caution. The implications of this behaviour are discussed further in § 6.

Epistemic and aleatoric predictive standard deviation of $b_{12}$

Figure 12. Epistemic (left) and aleatoric (right) predictive standard deviation of b12b_{12} for the withheld geometry. The epistemic component increases and concentrates where the geometry departs most from the training flows, while the aleatoric component follows the local correction magnitude and is more broadly distributed. The epistemic magnitude is nonetheless insufficient to account for the prediction error, consistent with the under-coverage in Figure 9.

Physical realisability of the predictions

We finally examine the physical realisability of the predicted anisotropy. In distribution, the vast majority of predicted states remain within the realisable barycentric region, with only a small fraction lying outside it, shown in Figure 13. The corresponding non-realisable fraction is 0.09% without the penalty and 0.58% with the penalty. The test RMSE\mathrm{RMSE} is essentially unchanged, from 0.0453 to 0.0451.

Figure 13. Physical realisability of the predicted anisotropy. In distribution, most predicted states remain within the realisable barycentric region, while the non-realisable fraction increases under geometric extrapolation. At the training budget considered here, the soft realisability penalty has no material effect on either the non-realisable fraction or the prediction error.

The same conclusion holds more strongly out of distribution. The fraction of non-realisable predictions increases to a few per cent, reaching 4.35% without the penalty and 5.08% with it, while the RMSE\mathrm{RMSE} changes from 0.1077 to 0.1102. At the training budget considered here, the soft penalty in (8) therefore does not materially improve realisability or predictive accuracy. Because the penalty targets the magnitude of the eigenvalue violations rather than the number of violating cells, and that fraction is already very small, it can even leave the fraction marginally higher.

These results also clarify the role of the tensor-basis representation. The basis enforces the required invariance, symmetry, and tracelessness of the anisotropy tensor, but it does not constrain its eigenvalues to the physically admissible range. The realisability penalty should therefore be viewed as a soft regularisation term rather than a guarantee. Enforcing realisability strictly would require a parameterisation that satisfies the eigenvalue or barycentric constraints by construction, as discussed further in § 6.

From a priori prediction to propagated uncertainty in wind-farm wakes

We next consider a three-dimensional engineering application in which uncertainty in the turbulence closure can directly affect quantities of practical interest: wind-farm wakes. Wake deficit and recovery determine the velocity available to downstream turbines and depend strongly on the turbulent-stress field, which can be difficult to represent with linear eddy-viscosity models. At the same time, wind-farm analysis often requires repeated flow evaluations, for which RANS remains a practical modelling approach. Quantifying the uncertainty associated with learned RANS corrections is therefore particularly relevant in this setting. This section first examines the a priori prediction of the learned three-dimensional correction and its transfer between wind-farm configurations, with the closure evaluated at the reference state at which it is defined, as Section 5.1 sets out. We then propagate posterior samples through the flow solver to determine how the closure uncertainty affects the predicted wake field.

Setup

We consider four atmospheric-boundary-layer cases: an isolated turbine and three six-turbine arrays with different spacings and aligned or staggered layouts. Matched high-fidelity reference data are available for all cases [10, 33]; the configurations and numerical resolutions are summarised in Appendix B.§ 3. The learned target is the Reynolds-stress anisotropy correction bijΔb_{ij}^{\Delta} relative to the baseline RANS model. Because these flows are fully three-dimensional, the correction is represented using the full ten-term tensor basis.

Unlike the anisotropy tensor considered in § 4, bijΔb_{ij}^{\Delta} is a signed correction and is not itself required to satisfy the realisability bounds imposed on bijb_{ij}. The penalty in (8) is therefore not applied to the wind-farm correction. Following [34], the training samples are restricted to the wake region using a production-based criterion. Cells are retained when the magnitude of the induced correction to turbulent-kinetic-energy production,

PΔ=−2kbijΔ∂jUi,(15)\mathcal{P}^{\Delta} = -2k b_{ij}^{\Delta} \partial_j U_i, \tag*{(15)}

exceeds a prescribed threshold. No additional geometric restriction, such as a wall-distance cutoff, is imposed. The wake-region criterion is used only to select the training data. At deployment, the trained closure is evaluated over the entire computational domain rather than being masked outside the wake. The predicted correction is comparatively small away from the wake but is retained as predicted instead of being forced to zero. This avoids introducing an abrupt discontinuity at the boundary of the training region, while allowing a small non-zero correction in the surrounding flow. Further details of the wake-region criterion are given in Appendix B.§ 3.

A further choice concerns the flow state at which the closure inputs are evaluated. The correction bijΔb_{ij}^{\Delta} is defined by a frozen inversion in which the mean flow is held at the high-fidelity reference, and the invariants that serve as closure inputs are evaluated on that same solution rather than on the uncorrected RANS field. This is the consistent choice for a constitutive relation: the correction is a function of the local state it describes, and pairing it with invariants taken from the uncorrected solution would instead learn a map contaminated by the errors of the baseline model, specific to that model and to the particular flow state it produces. The same convention is not used in § 4, where the target is the full anisotropy tensor and the inputs are the baseline RANS fields, following standard practice for that task; the distinction follows from the difference between learning a closure and learning a correction to one.

The consequence for deployment should be stated plainly. Because the inputs are evaluated on the reference state, the results reported below are conditional on that state being known, and they therefore measure the quality of the learned constitutive relation rather than end-to-end predictive performance on a configuration for which no high-fidelity data exist. On such a configuration the correction and the mean flow must be brought into agreement iteratively, by re-evaluating the closure as the solution evolves. The frozen propagation used in § 5.3 is accordingly a consistency and uncertainty-propagation study, bounded above by the frozen solution obtained with the reference correction, which is reported alongside it for exactly this reason. Coupled evaluation is left to future work.

A priori prediction and generalisation across layouts

When trained on one six-turbine array and evaluated on the other farm configurations, the closure transfers well across layouts. For the two withheld six-turbine cases, the wake-shear anisotropy correction is predicted with coefficients of determination of 0.87 and 0.85 and RMSE values of 0.03–0.04. These figures quantify how well the correction transfers as a function of the local state, evaluated at the reference state in both training and testing; they are not a measure of what a single uncorrected RANS solution would yield. Figure 14 shows a representative prediction, including the reference correction, BKAN posterior-mean field, pointwise error, and epistemic uncertainty. The main spatial structure of the correction is recovered, while the epistemic uncertainty is concentrated primarily around the high-shear wake regions.

A priori prediction of the three-dimensional Reynolds-stress anisotropy correction for a six-turbine array

Figure 14. A priori prediction of the three-dimensional Reynolds-stress anisotropy correction for a six-turbine array, shown on a streamwise–vertical plane through the wakes. The high-fidelity reference is compared with the BKAN posterior-mean prediction, pointwise absolute error, and epistemic standard deviation within the criterion-defined wake region. The main spatial structure of the correction is recovered, while the epistemic uncertainty is concentrated around the high-shear wake regions.

This transfer between related wind-farm configurations is more successful than the extrapolation between qualitatively different wall-bounded geometries considered in § 4.3. The two settings are not directly comparable in terms of their error metrics, since they involve different target fields and evaluation domains. Nevertheless, the results suggest that the local mapping for the wake correction transfers more readily between configurations that share the same underlying wake physics.

A more demanding test is obtained by training on the isolated single-turbine case and evaluating the model on the three multi-turbine arrays. In this setting, the RMSE increases to 0.05–0.07, with coefficients of determination between approximately 0.4 and 0.7, while the calibration error increases to 0.11–0.14. The isolated-turbine training case does not contain the wake–wake interactions and merged-wake structures present in the farm configurations, which provides a plausible explanation for the poorer transfer. For the subsequent solver propagation, we therefore use a model trained on a six-turbine array, where the training data already contain the multi-wake behaviour relevant to the deployment cases.

Because the predicted correction is subsequently propagated through the flow solver, we also examine the sensitivity of the deployed model to its architecture. Variations in network width, depth, and spline-grid resolution produce only small changes in the validation and test errors and in the predicted correction field. Across these configurations, the field changes by at most approximately 10% of its magnitude relative to the selected model. Removing the additional invariant features has the largest effect of the variations considered, giving both the highest errors and the largest change in the deployed correction field, although the absolute change remains modest. The complete sensitivity study is reported in Appendix B.§ 3. Based on this assessment, we fix the architecture used for the posterior ensemble propagated through the solver.

Propagating the posterior through the solver

The a priori uncertainty in bijΔb_{ij}^{\Delta} becomes practically relevant when it is propagated to quantities predicted by the flow solver. Moreover, a priori accuracy is not necessarily a reliable indicator of a posteriori performance [35], so the effect of the learned correction should also be assessed after propagation through the RANS equations.

We use a frozen-field propagation approach similar to that adopted for data-driven turbulence corrections in [8, 19]. Independent samples are drawn from the posterior distribution of the closure, and each sample is evaluated over the computational mesh to obtain a fixed correction field. The closure is trained only within the criterion-defined wake region but, as described in § 5.1, is evaluated over the full domain at deployment. The RANS solver is then run to convergence once for each fixed correction field. This differs from the coupled propagation strategy of [13], in which a stochastic algebraic model is embedded in the solver and its coefficient uncertainty is propagated using non-intrusive polynomial-chaos collocation.

During each RANS solve, the sampled correction field is held fixed and is not recomputed as the mean flow evolves. The ensemble of converged solutions then provides a distribution of the

propagated mean-flow quantities. We refer to predictions evaluated without the solver as a priori, while frozen propagation denotes the present procedure in which one converged RANS solution is obtained for each fixed sampled correction field. Fully coupled or online propagation, in which the learned correction is reevaluated during the nonlinear solver iterations, is not considered here.

Figure 15 and Figure 16 show vertical profiles of streamwise velocity (top row) and turbulence intensity (bottom row) at the wake centre, 4D4D downstream of the first four turbines. The comparisons include the baseline RANS solution, the propagated BKAN ensemble, a frozen “best-possible” solution obtained by injecting the correction derived from the LES reference, and the LES itself. The velocity is normalised by the inflow speed, U∞=8U_{\infty} = 8 m/s at the turbine hub height.

Figure 15. Aligned array (6T5D): vertical profiles at the wake centre, 4D4D downstream of the first four turbines. Top row, normalised streamwise velocity U/U∞U/U_\infty; bottom row, turbulence intensity. Baseline RANS (dash-dot), propagated BKAN ensemble (mean and shaded band), frozen best-possible solution (dashed), and LES reference (solid); the grey band marks the rotor-swept region.

Figure 16. Staggered array (6T7DSTG): vertical profiles at the wake centre, 4D4D downstream of the first four turbines. Rows and legend as in Figure 15.

For both unseen layouts, the baseline RANS predicts a weaker wake deficit than the LES across much of the rotor-swept region. Propagating the BKAN correction substantially improves the streamwise-velocity prediction. As summarised in Table 1, the rotor-region RMSE decreases from 0.119 to 0.034 for the aligned 6T5D array, corresponding to a reduction of approximately 72%, and from 0.170 to 0.015 for the staggered 6T7DSTG array, a reduction of approximately 91%. For the staggered case, the BKAN result approaches the frozen best-possible value of 0.012. Neither deployment layout is included in the training set, which contains only the aligned 6T7D array. The propagated results therefore test transfer to a different turbine spacing and, for 6T7DSTG, to a different arrangement.

LayoutBaselineBKANBest-possible
6T5D (aligned, unseen spacing)0.1190.034 (−72%)0.012
6T7DSTG (staggered, unseen layout)0.1700.015 (−91%)0.012

Table 1. Rotor-region streamwise-velocity root-mean-square error against LES, 4D downstream of each turbine for the two unseen wind-farm layouts. Results are shown for the baseline RANS solution, the propagated BKAN prediction, and the frozen best-possible solution. Percentages indicate the reduction relative to the baseline RANS error.

The propagated ensemble accounts for both epistemic and aleatoric contributions. For each realisation, the epistemic component is obtained from one posterior draw of the Bayesian closure parameters. The variance head is evaluated at the posterior-mean parameters to obtain the pointwise aleatoric variance σc2(λ)\sigma_c^2(\lambda) for each stored component cc of the symmetric correction tensor. Before propagation, the corresponding standard deviation σc\sigma_c is restricted to the criterion-defined wake region and clipped at its 99th percentile within that region. This mask applies only to the aleatoric perturbation; the BKAN correction itself remains evaluated over the entire computational domain.

Spatial correlation is imposed on the aleatoric component to represent the spatial structure of the model discrepancy. For each tensor component, an independent standard-normal field wc(x)w_c(\boldsymbol{x}) is generated over the structured mesh and passed through an anisotropic Gaussian filter. The resulting field is recentered and rescaled to unit global variance before being multiplied by the local predicted standard deviation,

ac(x)=σcuse(x)Gs[wc](x)−⟨Gs[wc]⟩std⁡(Gs[wc]),wc(x)∼N(0,1),(16)a_c(\boldsymbol{x})=\sigma_c^{\mathrm{use}}(\boldsymbol{x})\frac{\mathcal{G}_s[w_c](\boldsymbol{x})-\langle\mathcal{G}_s[w_c]\rangle}{\operatorname{std}(\mathcal{G}_s[w_c])}, \qquad w_c(\boldsymbol{x})\sim\mathcal{N}(0,1), \tag*{(16)}

where Gs\mathcal{G}_s denotes anisotropic Gaussian filtering and σcuse\sigma_c^{\mathrm{use}} is zero outside the wake region. Separate random fields are generated for the six stored tensor components; no cross-component correlation is imposed. Since a separate field is added per stored component, individual samples are not exactly trace-free; the isotropic part is dynamically inert in the incompressible solver (absorbed into the pressure) and does not affect the propagated statistics.

The spatial correlation scales are estimated once from the a priori residual field of the 6T7D training case. Directional autocorrelations of the residual are evaluated over the wake region, and the correlation length in each coordinate direction is defined by the first lag at which the autocorrelation falls below e−1e^{-1}. This gives

(Lx,Ly,Lz)=(600,50,40) m.(17)(L_x,L_y,L_z)=(600,50,40)\ \mathrm{m}. \tag*{(17)}

The same correlation lengths are used for all tensor components. For the Gaussian filter, they are converted to kernel standard deviations according to

sj=Lj2Δj(18)s_j = \frac{L_j}{2\Delta_j} \tag*{(18)}

where Δj\Delta_j is the grid spacing in direction jj. The correlation model is therefore determined entirely from the training case and does not use high-fidelity information from either deployment layout.

Each propagated realisation consists of one posterior draw of the BKAN parameters together with one independently generated correlated aleatoric field for each tensor component. Their sum is written once as a frozen correction field and held fixed during the corresponding RANS solve. In both deployment layouts, the LES velocity profile lies within the resulting propagated band over much of the rotor-swept region (Figures 15 and 16). The band is of a magnitude comparable to the observed BKAN–LES discrepancy, although these profile comparisons should not be interpreted as an independent calibration test.

Implications and limitations

The results show that the proposed closure can combine an inspectable constitutive representation with predictive uncertainty without sacrificing the accuracy obtained by the corresponding neural baseline. They also reveal important limits. In particular, the quality of the predicted uncertainty depends strongly on whether the deployment conditions are represented by the training data. This section discusses these findings, the choice of Bayesian parameterisation, and the main limitations of the present formulation.

Interpretability and Bayesian parameterisation

The comparison with the tensor-basis BNN shows that the interpretability of the KAN representation is obtained without a loss of predictive accuracy. In distribution, the BKAN and BNN achieve similar anisotropy errors, while the BKAN has the lower calibration error among the models compared in § 4.1. At the same time, the learned univariate edge functions can be inspected directly and, where their structure is sufficiently simple, approximated by closed-form expressions. The Bayesian treatment retains this property by placing distributions over the learned functions rather than replacing them with an opaque probabilistic mapping. This combination differs from the sparse Bayesian algebraic approach of [13]. There, uncertainty is introduced into an explicit closure selected from a prescribed dictionary of candidate terms. In the present approach, the scalar constitutive functions are instead learned directly through the KAN representation and can subsequently be inspected or distilled. This provides greater flexibility in the functional form, at the cost of a more complex learned model.

The parameterisation study further shows that the location of the posterior within the KAN matters. Treating the spline coefficients as Bayesian variables (b-coef) provides the most favourable overall balance of predictive accuracy, calibration, parameter cost, and interpretability. Placing the posterior only on the edge amplitudes (b-scale) gives poorer calibration, while treating both the spline coefficients and edge amplitudes as Bayesian variables (b-full) increases the epistemic spread and parameter count without improving calibration over b-coef. Although the detailed ordering of b-scale and b-full varies between experiments, b-coef is the preferred parameterisation in both the controlled regression study and the turbulence-anisotropy ablation (§ 3; Appendix § 12, Table Appendix C2). We therefore use b-coef as the default parameterisation throughout the turbulence applications.

Trust boundary under geometric extrapolation

The most important limitation concerns the behaviour of the predictive uncertainty under extrapolation. In the in-distribution tests considered here, the uncertainty is well calibrated and provides useful information about the model error: regions assigned larger uncertainty also tend to contain larger prediction errors. This behaviour changes when the closure is applied to a geometry that is absent from the training data. As shown in § 4.3, the epistemic uncertainty increases in regions affected by the geometric change, but not enough to account for the accompanying increase in prediction error. The result is substantial under-coverage and a marked deterioration in error ranking.

The model therefore retains some sensitivity to extrapolation without providing a quantitatively reliable measure of its consequences. An increase in epistemic uncertainty alone should not be interpreted as evidence that the predictive distribution remains calibrated outside the training regime. For the present closure, the in-distribution results support the use of the uncertainty estimate for interpolation within represented flow families, whereas uncertainty estimates for substantially different geometries should be interpreted cautiously.

Similar under-coverage under extrapolation has been reported for Bayesian turbulence corrections by [19], while [13] found that uncertainty intervals can widen in regions of larger discrepancy without necessarily capturing the full error. One likely contributor in the present model is the mean-field variational approximation, which neglects posterior correlations and may underestimate epistemic uncertainty when the model is taken away from the training distribution. Several extensions could address this limitation. Posterior tempering can increase the epistemic spread, although any improvement in coverage must be balanced against the resulting loss of in-distribution sharpness. Ensembles of independently trained BKANs could capture variation between different posterior modes that is not represented by a single mean-field approximation. More expressive variational families, such as low-rank or full-covariance Gaussian posteriors and normalising-flow approximations, could instead relax the posterior independence assumption directly. A complementary approach is to broaden the training set with flows that contain the geometric and physical variations expected at deployment. A systematic comparison of these strategies using the same calibration, sharpness, and sparsification measures employed here would be a useful direction for future work.

Realisability is regularised, not guaranteed

The eigenvalue penalty in (8) is a soft regularisation term rather than a hard constraint. The fraction of non-realisable predictions is already small in distribution without the penalty, and at the training budget used here the penalty does not materially reduce this fraction or improve the prediction error (Appendix § 12). The tensor-basis representation enforces invariance, symmetry, and tracelessness, but it does not constrain the eigenvalues of the predicted anisotropy to the realisable range.

A strict guarantee would therefore require a different parameterisation, for example one based on barycentric coordinates constrained to the admissible simplex or on a spectral representation that bounds the eigenvalues by construction. Such formulations are a natural extension of the present framework. The signed correction bijb_{ij} used for the wind-farm cases in § 5 is not itself subject to the same boundedness requirement.

Scope of the present study

Three aspects define the scope of the present results. First, the solver-level predictions are obtained using frozen propagation. Each sampled correction field is held fixed while the RANS equations converge, rather than being recomputed as the mean flow evolves. The results therefore demonstrate the effect of propagating the learned posterior through the solver, but do not address fully coupled a posteriori deployment.

Second, the symbolic distillation demonstrated here concerns individual, low-dimensional KAN functions. It provides a direct way to inspect and approximate learned constitutive relationships, but does not imply that the complete tensor-basis closure can always be reduced to a compact analytical expression. This distinction is particularly relevant where the tensor-basis coefficients are not uniquely identifiable.

Third, propagation of the aleatoric component requires a spatial correlation model in addition to the pointwise heteroscedastic variance predicted by the BKAN. In the present calculations, the correlation length is inferred from the training-case residual field. The sensitivity of the propagated uncertainty to this modelling choice has not been examined and should be considered in future solver-level uncertainty studies.

Concluding remarks

We have developed a Bayesian tensor-basis KAN for Reynolds-stress modelling that combines an inspectable constitutive representation with predictive uncertainty. Learnable spline functions on the KAN edges provide direct access to the scalar relationships underlying the closure and, where their structure is sufficiently simple, allow post hoc symbolic approximation. The variational treatment provides epistemic and aleatoric components of predictive uncertainty, while the tensor-basis formulation enforces Galilean and rotational invariance, symmetry, and tracelessness by construction.

The model was evaluated progressively from controlled regression to turbulent flows. On the regression benchmark, the BKAN achieves predictive performance comparable to the Bayesian neural baseline while providing well-calibrated uncertainty. Among the Bayesian KAN parameterisations considered, placing the posterior on the spline coefficients (b-coef) gives the most favourable overall balance of predictive accuracy, calibration, parameter cost, and interpretability. For Reynolds-stress anisotropy, the closure is accurate in distribution and its uncertainty is both well calibrated and informative: larger predicted uncertainty tends to coincide with larger prediction errors. This behaviour deteriorates under geometric extrapolation. Although the epistemic uncertainty responds to unfamiliar flow regions, it does not increase sufficiently to account for the corresponding prediction error, leading to substantial under-coverage and poorer error ranking. This defines a practical trust boundary for the present uncertainty model.

The three-dimensional wind-farm cases demonstrate that the same framework can be applied to an anisotropy correction and propagated through a RANS solver. Corrections learned from a six-turbine array transfer successfully to unseen wind-farm layouts, and frozen propagation substantially improves the resulting wake-velocity predictions. Relative to the baseline RANS solution, the rotor-region streamwise-velocity RMSE is reduced by approximately 72% for the unseen aligned 6T5D array and 91% for the unseen staggered 6T7DSTG array. These results show that the posterior can be propagated beyond the closure itself to quantify uncertainty in flow quantities of practical interest.

The main remaining challenges follow directly from the limitations identified in this study. Future work should focus on posterior approximations that better represent epistemic uncertainty under geometric shift, and fully coupled solver propagation in which the learned correction is updated as the flow evolves. These extensions would broaden the range over which the uncertainty estimate can be used reliably while retaining the interpretability provided by the KAN representation.

Acknowledgments

A.E. acknowledges support from the NWO Open Competition ENW-XS project U-TUBKAN (project number OCENW.XS25.3.099), funded by the Dutch Research Council (NWO).

CRediT authorship contribution statement

A. E.: Conceptualization (lead); Methodology (lead); Software & code (lead); Formal analysis (lead); Writing – original draft (lead). Kh. J.: Conceptualization (supporting); Methodology (supporting); Software & code (supporting); Formal analysis (equal); Writing – original draft (supporting). T. B.: Conceptualization (supporting); Methodology (supporting); Software & code (supporting); Formal analysis (supporting); Writing – original draft (supporting). R. P. D.: Conceptualization (equal); Methodology (equal); Supervision (lead); Writing – review and editing (equal).

Code and data availability

A reference implementation of the Bayesian Kolmogorov–Arnold network, together with the regression-benchmark and a priori anisotropy experiments, is openly available at https://github.com/alieid1st/bayesian-kan-turbulence under the MIT license. The UCI regression datasets follow the standard splits of [30], and the turbulence-anisotropy study uses the curated dataset of [32]; both are obtained from their original public sources as documented in the repository. The wind-farm large-eddy-simulation cases are not redistributed owing to their size and are available from the authors on reasonable request.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Model formulation and training

Variational formulation

We place a mean-field Gaussian posterior

qψ(θ)=∏iN(θi∣μi,σi2)(19)q_{\psi}(\boldsymbol{\theta}) = \prod_i \mathcal{N}(\theta_i \mid\mu_i, \sigma_i^2) \tag*{(19)}

over the parameters treated as Bayesian variables, as described in § 2.4. The standard deviation is parameterised as σi=softplus⁡(ρi)\sigma_i = \operatorname{softplus}(\rho_i) to ensure positivity. The model is trained by maximising the ELBO in (4).

The expected log-likelihood is estimated by Monte Carlo sampling. For the heteroscedastic Gaussian likelihood over the dd unique anisotropy components, the negative expected log-likelihood is approximated by

−Eq[log⁡p(D∣θ)]≈12∑pts∑c=1d[log⁡σc2(λ)+(bc−b^c)2σc2(λ)]+const.(20)-\mathbb{E}_{q}[\log p(\mathcal{D} \mid\theta)] \approx\frac{1}{2} \sum_{\mathrm{pts}} \sum_{c=1}^{d} \left[\log\sigma_c^2(\lambda) + \frac{(b_c-\hat{b}_c)^2}{\sigma_c^2(\lambda)}\right] + \mathrm{const}. \tag*{(20)}

For the Gaussian prior p(θi)=N(0,s2)p(\theta_i) = \mathcal{N}(0, s^2), the KL term has the closed-form expression

DKL(qψ∥p)=∑i[log⁡sσi+σi2+μi22s2−12].(21)D_{\mathrm{KL}}(q_{\psi}\Vert p) = \sum_i \left[\log\frac{s}{\sigma_i} + \frac{\sigma_i^2+\mu_i^2}{2s^2} - \frac{1}{2}\right]. \tag*{(21)}

To reduce gradient variance, the local reparameterisation trick is applied to the spline output of each edge. For the B-spline branch of the edge in (3), with Gaussian coefficients θk∼N(μk,σk2)\theta_k \sim\mathcal{N}(\mu_k,\sigma_k^2), the spline output at input xx is itself Gaussian,

ϕs(x)∼N(∑kBk(x)μk,∑kBk(x)2σk2),(22)\phi_s(x) \sim\mathcal{N}\left(\sum_k B_k(x)\mu_k,\sum_k B_k(x)^2\sigma_k^2\right), \tag*{(22)}

so the scalar spline output can be sampled directly rather than sampling every coefficient θk\theta_k individually. This reduces the variance of the gradient estimator during variational training. Edge-amplitude posteriors in the b-scale and b-full variants are sampled directly.

For the two-component scale-mixture prior

p(θ)=πN(0,s12)+(1−π)N(0,s22),(23)p(\theta) = \pi\mathcal{N}(0,s_1^2) + (1-\pi)\mathcal{N}(0,s_2^2), \tag*{(23)}

the KL term has no closed form and is estimated using the parameters sampled during the forward pass,

DKL≈∑i[log⁡qψ(θi)−log⁡p(θi)].(24)D_{\mathrm{KL}} \approx\sum_i \left[\log q_\psi(\theta_i) - \log p(\theta_i)\right]. \tag*{(24)}

The Gamma hyperprior p(s−2)=Gamma⁡(a0,b0)p(s^{-2}) = \operatorname{Gamma}(a_0,b_0) gives, through the evidence-maximising update of [26], the adaptive prior precision

λ⋆=a0+n/2b0+12Eq[∑iθi2],s2=1/λ⋆,(25)\lambda^\star= \frac{a_0+n/2}{b_0+\frac{1}{2}\mathbb{E}_q\left[\sum_i\theta_i^2\right]}, \qquad s^2 = 1/\lambda^\star, \tag*{(25)}

where nn is the number of parameters in the corresponding group. λ⋆\lambda^\star is recomputed during training and substituted into the Gaussian KL term. In all cases, the KL contribution is scaled by 1/N1/N and annealed from zero to one during training, as in (6). We found no material sensitivity of the results to the three prior formulations and therefore use the Gaussian prior unless stated otherwise.

At prediction time, we draw SS posterior samples from qψq_\psi, reconstruct the anisotropy through the fixed tensor-basis merge in (2) for each sample, and estimate the epistemic variance from the resulting ensemble. This is combined with the mean predicted aleatoric variance according to the law of total variance in (7). The reported point prediction is evaluated at the posterior-mean parameters. The complete training and prediction procedure is summarised in Algorithm 1.

Tensor basis, invariants and features

Let S^\widehat{\mathbf{S}} and W^\widehat{\mathbf{W}} denote the mean strain- and rotation-rate tensors normalised by a turbulent time scale. The integrity basis of [20] defines a set of tensor basis elements T(n)\mathbf{T}^{(n)} and scalar invariants λp\lambda_p. For two-dimensional mean flows, the following three-term basis is exact:

{S^, S^W^−W^S^, S^2−13tr⁡(S^2)I}(26)\left\{\widehat{\mathbf{S}},\ \widehat{\mathbf{S}}\widehat{\mathbf{W}}-\widehat{\mathbf{W}}\widehat{\mathbf{S}},\ \widehat{\mathbf{S}}^2-\frac{1}{3}\operatorname{tr}(\widehat{\mathbf{S}}^2)\mathbf{I}\right\} \tag*{(26)}

with invariants

λ1=tr⁡(S^2),λ2=tr⁡(W^2).(27)\lambda_1=\operatorname{tr}(\widehat{\mathbf{S}}^2), \qquad\lambda_2=\operatorname{tr}(\widehat{\mathbf{W}}^2). \tag*{(27)}

The normalised tensors are

S^=12τ(∇U+∇UT),W^=12τ(∇U−∇UT),(28)\widehat{\mathbf{S}}=\frac{1}{2}\tau\left(\nabla\mathbf{U}+\nabla\mathbf{U}^{\mathsf{T}}\right), \qquad\widehat{\mathbf{W}}=\frac{1}{2}\tau\left(\nabla\mathbf{U}-\nabla\mathbf{U}^{\mathsf{T}}\right), \tag*{(28)}

where τ\tau is the turbulent time scale, taken as 1/ω1/\omega for kk–ω\omega models and k/εk/\varepsilon for kk–ε\varepsilon models.

For three-dimensional flows, including the square duct and the turbine wakes, we use the full ten-term basis with five scalar invariants,

λ1=tr⁡(S^2),λ2=tr⁡(W^2),λ3=tr⁡(S^3),(29)\lambda_1=\operatorname{tr}(\widehat{\mathbf{S}}^2), \qquad\lambda_2=\operatorname{tr}(\widehat{\mathbf{W}}^2), \qquad\lambda_3=\operatorname{tr}(\widehat{\mathbf{S}}^3), \tag*{(29)}
λ4=tr⁡(W^2S^),λ5=tr⁡(W^2S^2).(30)\lambda_4=\operatorname{tr}(\widehat{\mathbf{W}}^2\widehat{\mathbf{S}}), \qquad\lambda_5=\operatorname{tr}(\widehat{\mathbf{W}}^2\widehat{\mathbf{S}}^2). \tag*{(30)}

The first six basis tensors are

T(1)=S^,(31)T^{(1)}=\widehat{S}, \tag*{(31)}
T(2)=S^W^−W^S^,(32)T^{(2)}=\widehat{S}\widehat{W}-\widehat{W}\widehat{S}, \tag*{(32)}
T(3)=S^2−13tr⁡(S^2)I,(33)T^{(3)}=\widehat{S}^{2}-\frac{1}{3}\operatorname{tr}(\widehat{S}^{2})I, \tag*{(33)}
T(4)=W^2−13tr⁡(W^2)I,(34)T^{(4)}=\widehat{W}^{2}-\frac{1}{3}\operatorname{tr}(\widehat{W}^{2})I, \tag*{(34)}
T(5)=W^S^2−S^2W^,(35)T^{(5)}=\widehat{W}\widehat{S}^{2}-\widehat{S}^{2}\widehat{W}, \tag*{(35)}
T(6)=W^2S^+S^W^2−23tr⁡(S^W^2)I.(36)T^{(6)}=\widehat{W}^{2}\widehat{S}+\widehat{S}\widehat{W}^{2}-\frac{2}{3}\operatorname{tr}(\widehat{S}\widehat{W}^{2})I. \tag*{(36)}

The tensors T(7)T^{(7)}–T(10)T^{(10)} are the remaining higher-order terms of the integrity basis given by [20]. Each basis tensor is unit-normalised before the merge step to improve numerical conditioning.

In addition to the tensor invariants, we use three bounded, Galilean-invariant scalar markers following [36, 37]: a strain–rotation balance qq, a turbulence-intensity measure TI, and a streamline pressure-gradient marker PG,

q=∥W^∥2−∥S^∥2∥W^∥2+∥S^∥2,TI=k12UiUi+k,(37)q=\frac{\lVert\widehat{W}\rVert^{2}-\lVert\widehat{S}\rVert^{2}}{\lVert\widehat{W}\rVert^{2}+\lVert\widehat{S}\rVert^{2}},\qquad \mathrm{TI}=\frac{k}{\frac{1}{2}U_iU_i+k}, \tag*{(37)}
PG=Uk∂kpUiUi∂jp∂jp+∣Uk∂kp∣.(38)\mathrm{PG}=\frac{U_k\partial_k p}{\sqrt{U_iU_i\partial_jp\partial_jp}+\lvert U_k\partial_kp\rvert}. \tag*{(38)}

The ratio forms make these quantities non-dimensional and bounded. They provide additional information about the local flow state, including turbulence level, the balance between strain and rotation, and acceleration or deceleration along the streamline. Together with the tensor invariants, these markers form the complete feature vector λ\lambda supplied to the BKAN. All scalar network inputs are standardised to zero mean and unit variance before training.

Architectures, training and reproducibility

The closure maps the invariant features to the tensor-basis coefficients using a KAN. Unless stated otherwise, the KAN has one hidden layer of width sixteen and uses cubic B-splines on a five-knot grid. A second KAN head, with the same architecture, predicts the heteroscedastic log-variance.

Training consists of two phases. First, the posterior means are deterministically pre-trained using a mean-squared-error objective. This short pre-training stage is needed for the network to resolve sharp fields. Second, the model is trained variationally by minimising the loss in (6). We use Adam optimizer with batch size 8192, learning rate 5×10−35\times10^{-3}, and linear KL annealing.

In-distribution metrics are averaged over three random seeds. A 15% validation hold-out is used to monitor over-fitting, and the corresponding train and validation convergence is shown in Figure Appendix A1. Predictive moments are computed using S=50S=50–100 posterior samples, following (7). Reported calibration metrics exclude the cross-plane shear components of planar flows, which are degenerate and identically close to zero. Including these components would artificially inflate the calibration metric.

The complete set of hyperparameters for each experiment is given in Table Appendix A1. The regression-benchmark width is matched to the reference BNN as described in Appendix B.§ 1; the turbulence and turbine models share a single architecture, with the realisability penalty active only on the bounded anisotropy target and disabled for the signed wake correction. No hyperparameter was tuned on the out-of-distribution or test splits; the validation hold-out was used only to confirm the absence of over-fitting.

SettingUCI benchmarkPlanar anisotropySquare ductTurbine wakes
Hidden layer(s)matched width[16][16][16]
Spline order / knotscubic / 5cubic / 5cubic / 5cubic / 5
Tensor-basis terms—31010
Invariants + markersdataset inputs2 + 35 + 35 + 3
Noise modelhomoscedasticheteroscedasticheteroscedasticheteroscedastic
Bayesian variantb-coefb-coefb-coefb-coef
PriorGaussianGaussianGaussianGaussian
Det. / VI epochs— / 40060 / 25060 / 25060 / 250
Batch size128819281928192
Learning rate10−210^{-2}5×10−35 \times10^{-3}5×10−35 \times10^{-3}5×10−35 \times10^{-3}
KL anneal (epochs)E/4E/4100100100
Realisability α\alpha010100
Predictive samples SS10050–10050–100100
OptimiserAdamAdamAdamAdam

Table A1. Hyperparameters and training protocol by experiment. “Det. / VI epochs” are the deterministic pre-training and variational-inference epoch counts; α\alpha is the realisability-penalty weight; SS is the number of posterior samples used for prediction.

Figure A1. Train and validation RMSE convergence for seed 0 in the in-distribution and out-of-distribution settings. In distribution, the train and validation curves remain close, indicating no evident over-fitting. The persistent out-of-distribution gap reflects the difficulty of geometric extrapolation.

Datasets and flow cases

Regression benchmark

The regression benchmark follows [30]. We use the standard datasets listed in Table Appendix B1, apply z-score standardisation, and use the standard random train–test splits. Errors and log-likelihoods are reported in the original target units.

For each dataset, the KAN width is chosen as the largest single-hidden-layer width for which the number of trainable parameters does not exceed that of the reference one-hidden-layer BNN. The comparison is therefore performed at a matched parameter budget. The Bayesian-parameterisation ablation shown in Figure 4 uses five of the benchmark datasets, with five random splits for each dataset.

DatasetddNNSplits
Boston housing1350620
Concrete8103020
Energy876820
Kin8nm8819220
Naval propulsion161193420
Power plant4956820
Wine (red)11159920
Yacht630820
Protein9457305

Table Appendix B1. Regression benchmark datasets following [30]: input dimension dd, number of instances NN, and number of standard splits.

Table Appendix B2 and Table Appendix B3 report the per-dataset test RMSE and test log-likelihood, averaged over the standard splits. The BKAN is competitive with the other models on both metrics across the suite; the corresponding calibration summary is shown in Figure 3.

DatasetBKANBNNMLPKANMC-drop.Deep ens.
Boston3.3673.1593.5044.5763.1343.144
Concrete4.9685.1775.1955.5275.1904.824
Energy0.3780.4980.4400.3411.3550.403
Kin8nm0.0880.0750.0790.0860.0960.074
Naval0.0000.0010.0010.0000.0010.000
Power3.8944.0153.9733.9054.1093.945
Wine0.6740.6330.6790.8470.6420.632
Yacht0.5510.8010.5770.5510.6970.480
Protein4.1244.4024.4724.1184.5854.405

Table Appendix B2. Per-dataset test RMSE (mean over the standard splits) on the regression benchmark; lower is better.

Turbulence-anisotropy cases

The Reynolds-stress anisotropy study in § 4 uses the curated dataset of [32], which provides matched baseline-RANS (kk–ω\omega SST) and high-fidelity (DNS/LES) fields interpolated onto common grids for a family of canonical flows. The cases used here are summarised in Table Appendix B4. The planar flows use the two-invariant three-term basis and the square duct the five-invariant ten-term basis.

DatasetBKANBNNMC-drop.Deep ens.
Boston−2.93−2.56−2.51−3.93
Concrete−3.15−3.07−2.97−2.91
Energy−0.47−0.77−1.90−1.04
Kin8nm1.011.170.991.30
Naval6.875.834.757.08
Power−2.78−2.81−2.82−2.76
Wine−1.17−0.96−2.09−1.35
Yacht−0.89−1.24−0.310.55
Protein−2.84−2.90−2.83−2.77

Table Appendix B3. Per-dataset test log-likelihood (mean over the standard splits) for the probabilistic models; higher is better.

In-distribution results use the parametric periodic-hill family, in which the hill slope is rescaled by a parameter α\alpha: the model is trained on α∈{0.5,0.8,1.2,1.5}\alpha\in\{0.5, 0.8, 1.2, 1.5\} and tested on the held-out α=1.0\alpha= 1.0 geometry at the same bulk Reynolds number. The out-of-distribution test instead withholds entire geometries: the model is trained on the periodic-hill and zero-pressure-gradient flat-plate flows and evaluated on the converging–diverging channel and the curved backward-facing step, neither of which appears in training. The square duct is treated separately, with training over a range of bulk Reynolds numbers and a held-out Reynolds number, to test the three-dimensional secondary-flow anisotropy that a linear eddy-viscosity model cannot represent. All in-distribution anisotropy metrics are reported as means over three random seeds.

CaseConfigurationRole
Periodic hills (α\alpha=0.5, 0.8, 1.2, 1.5)parametric slope, RebRe_b=5600, 2-Din-dist. train
Periodic hills (α\alpha=1.0)baseline slope, RebRe_b=5600, 2-Din-dist. test
Flat plate (ZPG boundary layer)Reθ≈Re_\theta\approx1000–4000, 2-DOOD train pool
Converging–diverging channelRe≈Re\approx12 600, 20 580, 2-DOOD test
Curved backward-facing stepRe≈Re\approx13 700, 2-DOOD test
Square ductRebRe_b=1100–3500, 3-Dsecondary flow

Table Appendix B4. Turbulence-anisotropy cases from the curated dataset of [32]. “In-dist.” denotes interpolation within a flow family and “OOD” a withheld geometry.

Wind-farm cases and wake-region criterion

The wind-farm dataset contains RANS solutions and matched high-fidelity references for the four cases summarised in Table Appendix B5. The training samples for the anisotropy correction are restricted to the wake region, while at deployment the trained closure is evaluated over the full computational domain, as described in § 5.1.

Following [34], the wake region is identified using a threshold on the magnitude of the induced correction to turbulent-kinetic-energy production,

∣PΔ∣>τP,PΔ=−2kbijΔ∂jUi,(39)\lvert\mathcal{P}^{\Delta}\rvert> \tau_{\mathrm{P}}, \qquad\mathcal{P}^{\Delta} = -2k b_{ij}^{\Delta}\partial_j U_i, \tag*{(39)}

where τP\tau_P denotes the prescribed production-correction threshold.

Figure Appendix B1 reports the architecture-sensitivity study used to select the configuration adopted for the wind-farm calculations. Variations in network width, depth, and spline-grid resolution produce only small changes in the validation and test RMSE. The RMS change in the deployed correction field is also small relative to the selected configuration. Removing the additional scalar features has the largest effect among the variants considered, giving both the highest validation and test errors and the largest change in the predicted correction field, although the absolute change remains modest.

CaseConfigurationMesh cells
1Tsingle turbine2.73×1062.73 \times10^{6}
6T5Dsix turbines, aligned, 5D5D2.73×1062.73 \times10^{6}
6T7Dsix turbines, aligned, 7D7D2.73×1062.73 \times10^{6}
6T7DSTGsix turbines, staggered, 7D7D2.73×1062.73 \times10^{6}

Table Appendix B5. Wind-farm cases. Each case is a full three-dimensional atmospheric-boundary-layer RANS solution with a matched high-fidelity reference.

!Architecture-sensitivity bar chart showing validation RMSE, test RMSE, and field change versus current for current [16], G5 full, wider [32], deeper [16, 16], finer grid [16], smaller [8], and no extra feats [16] bbox=0.24,0.46,0.75,0.70

Figure Appendix B1. Architecture sensitivity for the wind-farm correction model. Validation and test RMSE vary little with network width, depth, and spline-grid resolution. The RMS change in the deployed correction field is also small, with the largest change occurring when the additional scalar features are removed.

Extended results

Multi-seed results and Bayesian parameterisation

Table Appendix C1 reports the three-seed turbulence-anisotropy results for the b-coef parameterisation, while Table Appendix C2 compares the three Bayesian parameterisations on the periodic-hill in-distribution split. The multi-seed results confirm the contrast observed in the main text between the in-distribution and cross-geometry cases: both the prediction error and the calibration error increase substantially under geometric extrapolation.

For the parameterisation study, b-coef gives the lowest RMSE and calibration error among the three variants. The b-scale and b-full alternatives both produce larger epistemic uncertainty and less favourable predictive performance, although their detailed ordering differs. The mean epistemic uncertainty increases in the order coef < scale < full. Together with the regression benchmark in § 3, these results support the use of b-coef as the default Bayesian parameterisation.

SplitRegimeRMSECalibration error
Periodic hillsin-distribution0.046±0.0010.046 \pm0.0010.061±0.0130.061 \pm0.013
Cross-geometryout-of-distribution0.116±0.0090.116 \pm0.0090.255±0.0140.255 \pm0.014

Table C1. Three-seed turbulence-anisotropy results reported as mean ± standard deviation.

VariantParametersRMSECalibration errorMean epistemic
b-coef5 4720.0450.0670.0019
b-scale3 6480.0510.1080.0080
b-full6 0800.0490.1220.0085

Table C2. Bayesian-parameterisation ablation for turbulence anisotropy on the periodic-hill in-distribution split. The posterior is placed on the spline coefficients (b-coef), the edge amplitudes (b-scale), or both (b-full). Among the three variants, b-coef gives the lowest RMSE and calibration error, while the mean epistemic uncertainty increases in the order coef < scale < full.

Realisability penalty

Table Appendix C3 reports the realisability ablation underlying Figure 13. The fraction of non-realisable predictions is already small in distribution without the penalty. At the training budget considered here, introducing the soft penalty does not reduce this fraction and has little effect on the prediction error. Out of distribution, the fraction of non-realisable states increases to several per cent and is likewise not improved by the penalty.

The tensor-basis representation enforces invariance, symmetry, and tracelessness, but does not guarantee that the predicted eigenvalues remain within the realisable range. These results therefore support the use of a hard parameterisation if strict realisability is required, as discussed in § 6.

In-distributionOut-of-distribution
Non-realisableRMSENon-realisableRMSE
α=0\alpha= 00.09%0.04534.35%0.1077
α=10\alpha= 100.58%0.04515.08%0.1102

Table C3. Non-realisable fraction and test RMSE with the realisability penalty off (α=0\alpha= 0) and on (α=10\alpha= 10).

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the authors used ChatGPT (OpenAI) and Claude (Anthropic) for language editing, wording refinement, and code syntax checking. The authors reviewed and edited the output as needed and take full responsibility for the content of the published article.

References

  1. [1]A. Eidi, R. P. Dwight, A surrogate-based bayesian framework for parametric calibration of urans turbulence models informed by anisotropic pressure fluctuations, Flow, Turbulence and Combustion 117 (2026) 1.DOI
  2. [2]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.DOI
  3. [3]E. J. Parish, K. Duraisamy, A paradigm for data-driven predictive modeling using field inversion and machine learning, Journal of Computational physics 305 (2016) 758–774.DOI
  4. [4]J. Weatheritt, R. Sandberg, A novel evolutionary algorithm applied to algebraic modifications of the rans stress–strain relationship, Journal of Computational Physics 325 (2016) 22–37.DOI
  5. [5]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.DOI
  6. [6]K. Duraisamy, G. Iaccarino, H. Xiao, Turbulence modeling in the age of data, Annual Review of Fluid Mechanics 51 (2019) 357–377.DOI
  7. [7]H. Xiao, P. Cinnella, Quantification of model uncertainty in rans simulations: A review, Progress in Aerospace Sciences 108 (2019) 1–31.arxiv.org/abs/1806.10434
  8. [8]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.DOI
  9. [9]M. Emory, J. Larsson, G. Iaccarino, Modeling of structural uncertainties in reynolds-averaged navier-stokes closures, Physics of Fluids 25 (2013) 110822.DOI
  10. [10]A. Eidi, N. Zehtabiyan-Rezaie, R. Ghiassi, X. Yang, M. Abkar, Data-driven quantification of model-form uncertainty in reynolds-averaged simulations of wind farms, Physics of Fluids 34 (2022).DOI
  11. [11]R. McConkey, N. Kalia, E. Yee, F.-S. Lien, Realisability-informed machine learning for turbulence anisotropy mappings, Journal of Fluid Mechanics 1019 (2025) A49.
  12. [12]N. Kalia, R. McConkey, E. Yee, F.-S. Lien, Kolmogorov–arnold networks for turbulence anisotropy mapping, Physics of Fluids 37 (2025).arxiv.org/abs/2505.19366
  13. [13]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.DOI
  14. [14]Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljacic, T. Hou, M. Tegmark, Kan: Kolmogorov–arnold networks, in: International conference on learning representations, volume 2025, 2025, pp. 70367–70413.
  15. [15]S. A. Faroughi, F. Mostajeran, A. H. Mashhadzeh, S. Faroughi, Kolmogorov-arnold networks for data-driven, physics-informed, and deep-operator learning: a review, synthesis, and new analysis, Neural networks (2026) 108791.DOI
  16. [16]S. A. Faroughi, F. Mostajeran, A. Arzani, S. Faroughi, Symbolic–kan: Kolmogorov-arnold networks with discrete symbolic structure for interpretable learning, Journal of Computational Physics (2026) 115223.
  17. [17]M. M. Hassan, Bayesian kolmogorov-arnold networks: Uncertainty-aware interpretable modeling through probabilistic spline decomposition, Physica A: Statistical Mechanics and its Applications (2025) 131041.DOI
  18. [18]Z. Gao, G. E. Karniadakis, Scalable bayesian physics-informed kolmogorov-arnold networks, SIAM/ASA Journal on Uncertainty Quantification 13 (2025) 1543–1577.arxiv.org/abs/2501.08501
  19. [19]T. Buchanan, A. Eidi, R. P. Dwight, Bayesian neural network correction of rans turbulence models with uncertainty quantification in separated flows, arXiv preprint arXiv:2404.26034v (2024).
  20. [20]S. B. Pope, A more general effective-viscosity hypothesis, Journal of Fluid Mechanics 72 (1975) 331–340.DOI
  21. [21]Z. Liu, P. Ma, Y. Wang, W. Matusik, M. Tegmark, KAN 2.0: Kolmogorov–Arnold networks meet science, arXiv preprint arXiv:2408.10205 (2024).arxiv.org/abs/2408.10205
  22. [22]C. Blundell, J. Cornebise, K. Kavukcuoglu, D. Wierstra, Weight uncertainty in neural network, in: International conference on machine learning, PMLR, 2015, pp. 1613–1622.
  23. [23]D. P. Kingma, T. Salimans, M. Welling, Variational dropout and the local reparameterization trick, Advances in neural information processing systems 28 (2015).arxiv.org/abs/1506.02557
  24. [24]A. Kendall, Y. Gal, What uncertainties do we need in bayesian deep learning for computer vision?, Advances in neural information processing systems 30 (2017).arxiv.org/abs/1703.04977
  25. [25]D. A. Nix, A. S. Weigend, Estimating the mean and variance of the target probability distribution, in: Proceedings of 1994 ieee international conference on neural networks (ICNN’94), volume 1, IEEE, 1994, pp. 55–60.
  26. [26]D. J. MacKay, A practical bayesian framework for backpropagation networks, Neural computation 4 (1992) 448–472.DOI
  27. [27]S. Banerjee, R. Krahl, F. Durst, C. Zenger, Presentation of anisotropy properties of turbulence, invariants versus eigenvalue approaches, Journal of Turbulence 8 (2007) N32.DOI
  28. [28]Y. Gal, Z. Ghahramani, Dropout as a bayesian approximation: Representing model uncertainty in deep learning, in: international conference on machine learning, PMLR, 2016, pp. 1050–1059.arxiv.org/abs/1506.02142
  29. [29]B. Lakshminarayanan, A. Pritzl, C. Blundell, Simple and scalable predictive uncertainty estimation using deep ensembles, Advances in neural information processing systems 30 (2017).arxiv.org/abs/1612.01474
  30. [30]J. M. Hernández-Lobato, R. Adams, Probabilistic backpropagation for scalable learning of bayesian neural networks, in: International conference on machine learning, PMLR, 2015, pp. 1861–1869.
  31. [31]V. Kuleshov, N. Fenner, S. Ermon, Accurate uncertainties for deep learning using calibrated regression, in: International conference on machine learning, PMLR, 2018, pp. 2796–2804.arxiv.org/abs/1807.00263
  32. [32]R. McConkey, E. Yee, F.-S. Lien, A curated dataset for data-driven turbulence modelling, Scientific data 8 (2021) 255.DOI
  33. [33]A. Eidi, R. Ghiassi, X. Yang, M. Akbar, Model-form uncertainty quantification in rans simulations of wakes and power losses in wind farms, Renewable Energy 179 (2021) 2212–2223.DOI
  34. [34]K. Jigjid, A. Eidi, N. A. K. Doan, R. P. Dwight, Discovery of a physically interpretable data-driven wind-turbine wake model, Flow, Turbulence and Combustion 115 (2025) 1181–1207.
  35. [35]J. F. MacArt, J. Sirignano, J. B. Freund, Embedded training of neural-network subgrid-scale turbulence models, Physical Review Fluids 6 (2021) 050502.DOI
  36. [36]J.-X. Wang, J. Wu, J. Ling, G. Iaccarino, H. Xiao, A comprehensive physics-informed machine learning framework for predictive turbulence modeling, arXiv preprint arXiv:1701.07102 (2017).DOI
  37. [37]J.-L. Wu, H. Xiao, E. Paterson, Physics-informed machine learning approach for augmenting turbulence models: A comprehensive framework, Physical Review Fluids 3 (2018) 074602.

Paper details

vlnc:W_3pzjwr2t
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

ChatGPT (OpenAI), Claude (Anthropic)

View statement
Looking for related papers…
Contents