PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantifying and Reducing Model-Form Uncertainties in Reynolds-Averaged Navier-Stokes Simulations: A Data-Driven, Physics-Based Bayesian Approach

H. Xiao, J. -L. Wu, J. -X. Wang, R. Sun, C. J. Roy

arXiv:1508.06315v3physics.comp-phphysics.flu-dyn

TL;DR

RANS turbulence models remain widely used despite substantial model-form uncertainty, motivating methods that can quantify errors without the cost of high-fidelity simulations. This paper develops an open-box, physics-informed Bayesian framework that perturbs Reynolds stresses and assimilates sparse observations through iterative ensemble Kalman inference. Across periodic-hill and square-duct flows, posterior velocities show significantly better agreement with benchmark data than baseline results, while the framework remains subject to limitations in sparse-data settings.

  • Problem

    RANS turbulence models are widely used but their model-form uncertainties are difficult to quantify, especially when observations are sparse.

  • Method

    The framework introduces compact, physically constrained uncertainties directly into Reynolds stresses and combines prior turbulence knowledge with data using approximate Bayesian ensemble inference.

  • Results

    Posterior mean velocities show significantly better agreement with benchmark data than baseline results, using periodic-hill and square-duct cases.

  • Takeaways & Limitations

    The approach provides an open-box, physics-informed framework for quantifying and reducing model-form uncertainties in RANS simulations.

  • Takeaways & Limitations

    In a sparse-observation setting, the available velocity data may be insufficient to constrain all uncertainties.

Abstract

from arXiv · show

Despite their well-known limitations, Reynolds-Averaged Navier-Stokes (RANS) models are still the workhorse tools for turbulent flow simulations in today's engineering application. For many practical flows, the turbulence models are by far the largest source of uncertainty. In this work we develop an open-box, physics-informed Bayesian framework for quantifying model-form uncertainties in RANS simulations. Uncertainties are introduced directly to the Reynolds stresses and are represented with compact parameterization accounting for empirical prior knowledge and physical constraints (e.g., realizability, smoothness, and symmetry). An iterative ensemble Kalman method is used to assimilate the prior knowledge and observation data in a Bayesian framework, and to propagate them to posterior distributions of velocities and other Quantities of Interest (QoIs). We use two representative cases, the flow over periodic hills and the flow in a square duct, to evaluate the performance of the proposed framework. Simulation results suggest that, even with very sparse observations, the posterior mean velocities and other QoIs have significantly better agreement with the benchmark data compared to the baseline results. At most locations the posterior distribution adequately captures the true model error within the developed model form uncertainty bounds. The framework is a major improvement over existing black-box, physics-neutral methods for model-form uncertainty quantification, where prior knowledge and details of the models are not exploited. This approach has potential implications in many fields in which the governing equations are well understood but the model uncertainty comes from unresolved physical processes.

1. Introduction

RANS remains widely used because high-fidelity simulations are expensive, but turbulence-model discrepancies are a major uncertainty source. Existing approaches motivate an open-box framework that injects uncertainty into modeled Reynolds stresses while using turbulence-model knowledge.

  • Motivation: RANS remains a workhorse for engineering turbulent-flow simulations despite poor performance in separation, mean pressure-gradient, and mean-flow-curvature regimes.LES and DNS are often too expensive for engineering design and optimization workflows requiring short turnaround times.
  • Model-form uncertainty: Turbulence closure models are the main focus because empirical models cannot accurately represent regime-dependent, physics-rich turbulent phenomena.Different models perform better in different cases, and none is generally superior.
  • Existing approaches: Black-box discrepancy methods do not exploit prior information about model-form uncertainty and often represent errors only at the level of quantities of interest.RANS modeling errors arise specifically from the modeled Reynolds stress term.
  • Existing approaches: Open-box approaches inject uncertainty into closure-model quantities, including Reynolds stresses, rather than perturbing numerical-model outputs directly.Related methods include Reynolds-stress discrepancies, limiting-state perturbations, and eddy-viscosity adjustments.
  • Contribution: The proposed framework combines statistical inference with turbulence-model domain knowledge to quantify and reduce RANS model-form uncertainty.Its design emphasizes physical realizability and spatial correlations while using an ensemble-based approach.

2.1. Prior Knowledge in RANS Modeling

The framework encodes prior knowledge that RANS uncertainty is concentrated in modeled Reynolds stresses and should respect physical, spatial, and problem-specific structure.

  • Prior knowledge: Modeled Reynolds stresses are identified as the main source of uncertainty in RANS predictions.This follows from the composite nature of RANS equations, which combine reliable conservation laws with approximate Reynolds-stress models.
  • Physical realizability: Physically realizable Reynolds stresses occupy a constrained subspace, motivating parameterizations that remain within the realizable set.The true stress at each point resides in a subspace of a six-dimensional space.
  • Spatial and problem-specific structure: Reynolds-stress fields usually vary smoothly in space, so uncertainty representations should preserve spatial smoothness.The framework also uses problem-specific knowledge about where eddy-viscosity models are expected to perform poorly.
  • Spatial and problem-specific structure: For periodic-hill flow, larger Reynolds-stress discrepancies are expected in recirculation, nonparallel free-shear, and strong mean-flow-curvature regions.These regional expectations are represented through a spatial variance field.

2.2. Representations of Prior Knowledge in the Modeling Framework

The framework represents Reynolds-stress uncertainty through physically meaningful variables and compact, smooth random-field parameterizations. Transformations and bounded coordinates encode realizability while basis choices encode spatial and problem-specific prior knowledge.

  • Physical representation: Reynolds stresses are transformed into six physically meaningful variables: turbulent kinetic energy, anisotropy shape, and tensor orientation.The variables are denoted (k, ξ, η, v1, v2, v3), with k non-negative and the orientation represented by orthonormal vectors.
  • Perturbation choices: Uncertainty is injected into k, ξ, and η while tensor orientations are left unperturbed because orientation perturbations can destabilize the RANS momentum equation.The logarithmic discrepancy of k also preserves the non-negativity of turbulent kinetic energy.
  • Limitations: Bounding natural coordinates enforces realizability but can distort the prior distribution, especially when baseline stresses lie near realizability boundaries.The resulting distribution may become non-Gaussian or concentrated near boundaries.
  • Problem-specific prior: Problem-specific prior knowledge is encoded through the choice of spatial basis and variance field.For periodic hills, the variance reflects regions expected to have larger modeled Reynolds-stress discrepancies.
  • Random-field parameterization: The discrepancies are modeled as zero-mean Gaussian fields with spatial covariance and then projected onto deterministic basis functions.The basis representation reduces uncertainty dimension and ensures spatial smoothness; orthogonality is not required.

2.3. Inverse modeling based on an iterative ensemble Kalman method

An iterative ensemble Kalman inversion augments the physical state with discrepancy coefficients, repeatedly solves the RANS equations, and assimilates velocity observations. The converged ensemble represents the posterior state distribution, while sparse observations and approximate inference constrain accuracy.

  • State augmentation: The state is augmented to include both the velocity field and unknown discrepancy coefficients, x ≡ [u, ω]^T.This allows the inference procedure to update physical variables and model-discrepancy parameters together.
  • Iterative inference: Each iteration reconstructs Reynolds stresses from coefficients, solves the RANS equations for velocities, and assimilates velocity observations with Kalman filtering.The updated ensemble is then used in subsequent iterations.
  • Posterior construction: The iterations stop when the prediction–observation misfit falls below the observation noise level.The converged ensemble is treated as a sample-based posterior distribution from which moments can be computed.
  • Observation model: Process noise accounts for differences between the observed system and the numerical model and allows prior and likelihood supports to overlap.Posterior means are reported as insensitive to observation-noise level when σobs exceeds 1% of truth.
  • Inference properties: The method searches the linear space spanned by the prior ensemble for a solution that reduces observation misfit.Its ensemble Kalman formulation is computationally cheaper than exact Bayesian inference but only approximate.
  • Limitations: Sparse observations may be insufficient to constrain all state uncertainties, and the ensemble Kalman approximation is not expected to match exact Bayesian posterior accuracy.The forward model provides physical regularization for the ill-posed inverse problem.

2.4. Summary of the algorithm in the proposed framework

The framework begins with a baseline RANS solution, parameterizes Reynolds-stress discrepancies using KL modes, and iteratively assimilates velocity observations to obtain a posterior state ensemble.

  • The algorithm first performs a baseline RANS simulation to obtain velocity and Reynolds-stress fields.
  • KL expansion constructs a basis set, which is used to generate an initial prior ensemble of coefficient vectors.
  • Each iteration recovers discrepancy fields from the coefficient ensemble and basis functions, then obtains Reynolds-stress realizations and corresponding velocity fields.
  • The ensemble mean is compared with velocity observations, and Kalman filtering updates the augmented state ensemble x_j = [u_j, ω_j]^T.
  • Iterations stop when statistical convergence of the ensemble is achieved.

3. Implementation and Numerical Methods

The implementation combines OpenFOAM-based baseline and forward RANS solvers with KL expansions and an iterative ensemble Kalman method, reducing the cost of repeated forward simulations.

  • The framework combines RANS models, Reynolds-stress mapping, KL expansions, and iterative ensemble Kalman inference.
  • The baseline simpleFoam solver computes velocity and Reynolds stresses by solving the RANS and turbulence-quantity equations.
  • The Launder–Sharma low Reynolds number k–ε model is used for baseline simulations, with wall-normal mesh refinement to resolve the boundary layer without wall functions.
  • The forward tauFoam solver computes velocity directly from a given Reynolds stress field, without specifying or solving a turbulence model.
  • Forward simulations are initialized from converged baseline solutions, requiring fewer iterations than baseline simulations.
  • 10% of the computational cost as that of the baseline simulation is required by forward RANS simulations to achieve the same residual.

4. Numerical Simulations

Across periodic-hill and square-duct simulations, assimilating sparse velocity observations generally improves posterior velocities and other QoIs while reducing uncertainty, but performance depends on observation coverage and model parameterization.

  • Periodic hills: The framework significantly improves posterior velocity predictions relative to baseline results in the periodic-hill case.
  • Periodic hills: The remaining velocity differences are attributed to sparse observations and inadequacy of the inference model, including non-unique Reynolds-stress-to-velocity mapping.
  • Scope and limitations: The method introduces uncertainty only in Reynolds-stress magnitude and shape, retains limited KL modes, and therefore represents smooth discrepancy fields without varying orientations.
  • Periodic hills: 95% posterior credible intervals are significantly narrower than prior intervals and better cover benchmark data where observations are available.
  • Periodic hills: Between x/H = 1 and x/H = 8, posterior samples improve agreement with DNS for both wall shear stress and reattachment point.
  • Square duct: In the square duct, posterior velocities improve along all four cross sections, with reduced scattering while covering the truth in most regions.
  • Square duct: The posterior mean agrees well with benchmark secondary-flow direction, intensity, and vortex-center location at most locations.

5. Discussion

The framework improves velocity and QoI predictions with sparse observations, while its computational cost, correlation assumptions, and Reynolds-stress identifiability constrain its use. The discussion also reports feasibility, posterior uncertainty reduction, and limitations of the approximate inference method.

  • Computational cost: Parallel execution on 60 CPU cores makes the uncertainty quantification procedure's wall time approximately equal to the baseline simulation.The 60 forward simulations are run simultaneously in each iteration, assuming the baseline simulation is run on a single core.
  • Correlation structure and experimental design: In three-dimensional complex flows, more measurements may be needed, especially when distinct regions are weakly correlated.Experimental design should ensure that regions of interest have measurement data where possible.
  • Correlation structure and experimental design: Correlation length scales must reflect the mean flow: overly small scales fail to correct unobserved regions, whereas overly large scales produce spurious corrections.Correlation structure can be studied a priori from an ensemble of RANS simulations, and observations can subsequently be optimized using correlation studies.
  • Framework performance: Even with velocity observations at very few locations, posterior velocities are significantly improved compared to baseline results.The numerical simulations demonstrate the feasibility of the framework, and the posterior ensemble also quantifies and reduces modeled Reynolds-stress uncertainties.
  • Framework limitations: Velocity observations generally constrain a lower-dimensional observable Reynolds-stress projection rather than the full Reynolds-stress field.The mapping from Reynolds stresses to mean velocities is many-to-one, while the relevant projection is observable from a control-theoretic perspective.
  • Framework limitations: The iterative ensemble method is computationally affordable but approximate, so its posterior distribution may deviate from the true distribution.More accurate Markov Chain Monte Carlo methods are described as prohibitively expensive for this setting.

6. Conclusion

The paper proposes a physics-informed Bayesian framework that introduces uncertainty directly into Reynolds stresses and assimilates prior knowledge with observations. In two test cases, sparse observations improved posterior velocities and QoIs relative to baseline results, while important limitations remained in Reynolds-stress accuracy and inference approximation.

  • The framework quantifies RANS model-form uncertainty by introducing uncertainty directly into Reynolds stresses through compact parameterization.The parameterization incorporates empirical prior knowledge and physical constraints including realizability, smoothness, and symmetry.
  • An iterative ensemble Kalman method assimilates prior knowledge and observations in a Bayesian framework and propagates uncertainty to posterior Reynolds stresses and QoIs.The method was demonstrated on periodic-hill and square-duct flows.
  • Even with sparse observations, posterior mean velocities show significantly better agreement with benchmark data than baseline results.
  • The inferred full Reynolds-stress field is not accurate because the uncertainty space is high-dimensional, velocity observations are sparse, and the stress-to-velocity mapping may be non-unique.The authors nevertheless argue that the inferred stresses remain valuable and can be extrapolated to cases with similar physical characteristics.
  • The iterative ensemble method is computationally less intensive but less accurate than exact Bayesian inference based on Markov Chain Monte-Carlo sampling.The impact of this approximate Bayesian inference method is left for future study.

Appendix A. Mapping from Barycentric Coordinates to Natural Coordinates

The appendix maps Reynolds-stress representations between Barycentric, Cartesian, and natural coordinates to support physically interpretable uncertainty parameterization. The transformed variables are perturbed as random fields and mapped back to Reynolds stresses.

  • Reynolds-stress anisotropy eigenvalues are linearly transformed into Barycentric coordinates C1, C2, and C3.The coordinates represent portions of the three sub-triangles and sum to 1.
  • Points inside the Barycentric triangle are represented as convex combinations of its three vertices in Cartesian coordinates.
  • Natural coordinates are used to parameterize the entire triangle while minimizing artificial capping of Reynolds stresses outside the realizable range.The transformation uses standard finite-element shape functions.
  • The inverse Cartesian-to-natural mapping is nontrivial because it requires solving a bilinear equation system, so analytical results are used.
  • Baseline RANS Reynolds stresses are mapped to turbulent kinetic energy and anisotropy-shape variables, which are modeled as random fields and mapped back after perturbation.

Appendix B. Iterative Ensemble Kalman Method for Inverse Modeling

The iterative ensemble Kalman method repeatedly predicts RANS velocities from an ensemble of uncertain states and updates that ensemble using observations. Iterations continue until the ensemble is statistically converged.

  • The method begins with a baseline RANS velocity prediction, observations, and an observation-error covariance matrix.
  • An initial ensemble of augmented states is generated in the sampling step.
  • Each ensemble state is propagated with the forward model by reconstructing Reynolds-stress fields and computing velocities from the RANS equations.
  • The analysis step computes ensemble statistics and the Kalman gain, then updates each predicted sample using the observations.The observation vector is mapped from state space to observation space by H.
  • Prediction and analysis are repeated until the ensemble is statistically converged.
Loading 1508.06315v3…