Physics-Guided Machine Learning for Uncertainty Quantification in Turbulence Models

Conference: NEURIPS2025 arXiv: 2511.05633 Code: To be confirmed Area: Scientific Computing Keywords: turbulence modeling, uncertainty quantification, CNN, eigenspace perturbation, RANS, physics-guided ML

TL;DR

This paper proposes a hybrid ML–EPM framework that employs a lightweight CNN to learn a correction mapping from RANS turbulent kinetic energy fields to DNS ground truth, using the learned corrections to modulate the perturbation magnitude of the Eigenspace Perturbation Method (EPM). The approach reduces uncertainty estimation errors by 1–2 orders of magnitude while preserving physical consistency.

Background & Motivation

Background: Turbulence is the dominant mechanism for momentum, heat, and mass transfer in natural and engineering systems. Direct numerical simulation (DNS) is computationally intractable for high-Reynolds-number flows, so engineering practice relies on turbulence models (e.g., k-ε, k-ω eddy-viscosity models) to approximate the effects of unresolved scales.

Limitations of Prior Work: The Reynolds-averaging procedure introduces an unknown Reynolds stress tensor. Turbulence models close the governing equations through empirical constitutive relations, but these assumptions (e.g., the Boussinesq hypothesis) break down significantly in complex flows involving curvature, adverse pressure gradients, and separation, introducing substantial epistemic uncertainty.

Key Challenge: Deterministic CFD provides only a single prediction that may carry large errors. Uncertainty quantification (UQ) yields probabilistic predictions, which are critical for estimating safety margins in engineering design and for distinguishing physical phenomena from model artifacts in scientific research. The EPM, the de facto standard for turbulence model UQ, explores model uncertainty by applying physically constrained perturbations to the eigenvalues, eigenvectors, and magnitude of the Reynolds stress ellipsoid. However, EPM determines maximum allowable perturbations purely from physical principles, leading to overly wide uncertainty bounds and imprecise calibration.

Goal: Different flow configurations and different spatial regions within the same flow exhibit varying magnitudes of turbulence model error. EPM requires a mechanism to modulate perturbation magnitude according to local flow conditions—a capability that physical principles alone cannot provide. Neural networks can learn error mappings from paired high-fidelity (DNS) and turbulence model data, enabling a hybrid framework in which physics governs how to perturb and ML governs how much to perturb.

Method

Overall Architecture

The hybrid ML–EPM framework operates on two levels. At the physics level, EPM prescribes the direction and structure of perturbations while preserving the anisotropy tensor \(b_{ij}\). At the data-driven level, a CNN learns a correction mapping from the RANS-predicted turbulent kinetic energy (TKE) field to the DNS ground truth; the corrected TKE replaces the original RANS prediction to modulate the perturbation magnitude.

Key Designs

Module 1: Eigenspace Decomposition and Perturbation of Reynolds Stresses

• Function: Decomposes the Reynolds stress tensor \(R_{ij}\) into the product of TKE \(k\), eigenvalue matrix \(\Lambda\), and eigenvector matrix \(v\), and applies physically constrained perturbations in this eigenspace.
• Mechanism: \(R_{ij} = 2\rho k(v \cdot \Lambda \cdot v^\top + \tfrac{1}{3}\delta_{ij})\); after perturbation, \(R^*_{ij} = 2\rho k^*(v^* \cdot \Lambda^* \cdot v^{*\top} + \tfrac{1}{3}\delta_{ij})\). This work applies data-driven correction only to \(k\), leaving the RANS-predicted anisotropic structure (\(v_{ij}\), \(\Lambda_{ij}\)) unchanged.
• Design Motivation: Modifying only the TKE magnitude while retaining the anisotropy direction prevents the non-physical stress distortions common in purely data-driven approaches and ensures that the corrected Reynolds stresses satisfy realizability conditions.

Module 2: CNN for TKE Correction Mapping

• Function: Corrects the RANS-predicted TKE field \(k^{\text{RANS}}(x,y)\) toward the DNS TKE field \(k^{\text{DNS}}(x,y)\).
• Mechanism: The TKE correction is framed as a supervised learning task \(\hat{g}(k^{\text{RANS}};\theta) \to k^{\text{DNS}}\). A lightweight 1D-CNN is employed: 2 convolutional layers (kernel size 3) + max pooling + 2 fully connected layers, totaling approximately 86 trainable parameters. ReLU activations and Batch Normalization follow each convolutional layer.
• Design Motivation: With limited training data (only two canonical flow configurations), an extremely compact network (86 parameters) prevents overfitting while maintaining interpretability. The 1D-CNN processes TKE profiles along the wall-normal direction, naturally suited to capturing spatially local features.

Module 3: Injection of Corrected TKE into EPM

• Function: Replaces the original RANS TKE with the CNN-predicted corrected TKE to reconstruct the modified Reynolds stress tensor.
• Mechanism: \(R^{\text{corr}}_{ij} = 2\hat{k}^{\text{DNS}} \cdot b^{\text{RANS}}_{ij}\), where \(b^{\text{RANS}}_{ij}\) is the normalized anisotropy tensor from RANS.
• Design Motivation: This coupling achieves a clear division of labor—physics governs structure and ML governs magnitude—leveraging EPM's physical consistency guarantees while overcoming the inability of purely physics-based methods to precisely calibrate perturbation magnitude.

Loss & Training

• Training employs Mean Absolute Error (MAE) loss: \(\mathcal{L}(\theta) = \frac{1}{N}\sum_i |\hat{g}(k_i^{\text{RANS}}) - k_i^{\text{DNS}}|\)
• Optimizer: Adam, learning rate \(10^{-3}\)
• Early stopping: patience = 10 epochs to prevent overfitting
• Data split: 75% training / 5% validation / 20% test

Key Experimental Results

Main Results

Table 1: Error Improvement at Chordwise Locations for the SD7003 Airfoil

Chordwise Location	RANS MAE Order	CNN-Corrected MAE Order	Error Reduction
Leading region (attached flow)	~\(10^{-2}\)	~\(10^{-4}\)	~2 orders of magnitude
Mid-chord (transition zone)	~\(10^{-2}\)	~\(10^{-3}\)–\(10^{-4}\)	1–2 orders of magnitude
Trailing region (wake)	~\(10^{-2}\)	~\(10^{-3}\)–\(10^{-4}\)	1–2 orders of magnitude

Table 2: Error Improvement at Streamwise Locations for the Periodic Hill

Streamwise Position \(x/h\)	Flow Feature	RANS vs. DNS	CNN-Corrected vs. DNS	Improvement
2.057	Inside separation bubble	Large deviation	Significant improvement	~2 orders of magnitude
4.769	Near reattachment point	Evident error	Approaches DNS	~2–3 orders of magnitude
5.342	Downstream of reattachment	Noticeable deviation	Nearly coincides with DNS	~3 orders of magnitude
Downstream developed region	Fully developed boundary layer	Smaller error	Further reduced	~1–2 orders of magnitude

Key Findings

1. 1–2 orders of magnitude error reduction: For both the SD7003 airfoil and periodic hill benchmarks, the MAE of CNN-corrected TKE profiles relative to DNS is approximately 1–2 orders of magnitude lower than the baseline RANS (up to 3 orders at select locations).
2. Greatest improvement in separated flow regions: CNN correction is most pronounced in flow separation and reattachment zones—where RANS model errors are largest—with corrected profiles nearly coinciding with DNS.
3. No data leakage: The periodic hill test case was excluded from training, validating the model's generalization capability.
4. Effectiveness of a minimal architecture: A 1D-CNN with only 86 trainable parameters achieves substantial improvement, indicating that the TKE correction mapping has limited complexity and does not require large models.
5. Physical consistency: Retaining the RANS anisotropic structure avoids non-physical stress distortions, achieving error reductions comparable to purely data-driven methods (e.g., field-inversion neural networks) while maintaining full physical interpretability.

Highlights & Insights

1. Elegant division of labor between "how to perturb" and "how much to perturb": Physical principles determine the direction and structure of perturbations; ML determines the magnitude—fully exploiting the complementary strengths of both.
2. Extreme parsimony: An 86-parameter network is a well-reasoned design choice under limited data, reflecting a deep understanding of the problem's intrinsic complexity.
3. Natural extension of the EPM framework: The method integrates seamlessly into existing EPM workflows without requiring a redesign of the entire UQ pipeline, facilitating industrial adoption.
4. Preserved physical interpretability: Modifying only a scalar field (TKE) rather than a tensor field guarantees realizability and interpretability of the correction.

Limitations & Future Work

1. Limited dataset: Validation is restricted to two 2D canonical flows (SD7003 airfoil and periodic hill); 3D flows and higher-Reynolds-number configurations remain untested.
2. DNS data dependency: Paired RANS–DNS data are required for training, and DNS is prohibitively expensive for complex engineering flows.
3. TKE-only correction: Data-driven correction is not applied to the anisotropic structure (eigenvalues and eigenvectors), which may be insufficient for strongly anisotropic flows.
4. Limited scope as a workshop paper: More detailed ablation studies and statistical analyses are absent.
5. Generalization boundaries unclear: Model performance outside the training flow types (e.g., rotating flows, compressible flows) is unknown.

Related Work & Insights

• Original EPM framework (Iaccarino et al., Mishra et al.): A purely physics-based method; the present work introduces ML-based modulation on top of this foundation.
• Heyse et al.: An earlier exploration of ML within EPM; this paper further simplifies and validates the CNN approach.
• Parish & Duraisamy (field-inversion): A purely data-driven field-inversion approach achieving comparable error reduction but lacking physical interpretability.
• Key Insight: The physics+ML hybrid paradigm has broad applicability—any scientific computing problem with a physical model requiring calibration can potentially adopt a similar data-driven correction strategy.

Rating

• Novelty: ⭐⭐⭐ (The idea is well-articulated, but the core contribution is embedding CNN into an existing EPM framework, which is conceptually straightforward)
• Experimental Thoroughness: ⭐⭐⭐ (Only two 2D cases, limited by the workshop paper format)
• Writing Quality: ⭐⭐⭐⭐ (Clear exposition of physics and methodology; motivation and positioning are accurate)
• Value: ⭐⭐⭐⭐ (Provides a practical and deployable improvement path for turbulence model UQ)

Related Papers

• [AAAI 2026] Scientific Knowledge-Guided Machine Learning for Vessel Power Prediction: A Comparative Study
• [NeurIPS 2025] F-Adapter: Frequency-Adaptive Parameter-Efficient Fine-Tuning in Scientific Machine Learning
• [NeurIPS 2025] Neural Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Data
• [NeurIPS 2025] GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations
• [NeurIPS 2025] EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale