🔬 physics

Goal-Oriented Weighting of Reynolds-Stress Data for Learning Turbulence Models in Complex Flows

This paper proposes a goal-oriented weighting strategy that leverages adjoint sensitivities to prioritize specific Reynolds-stress components during the supervised training of data-driven turbulence models, thereby improving predictions of downstream quantities of interest in complex flows without requiring additional expensive RANS simulations.

Original authors: Zhuolin Zhao, Haochen Wang, Youngwoo Kim, Solkeun Jee, Heng Xiao

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhuolin Zhao, Haochen Wang, Youngwoo Kim, Solkeun Jee, Heng Xiao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict how air moves over a jet engine or how coolant spreads across a turbine blade. Engineers rely on computer simulations to answer these questions, but the air in these machines is turbulent, churning in chaotic swirls that are impossible to track perfectly. To make the math manageable, scientists use a simplified approach that averages out the chaos, but this averaging leaves a gap: a missing piece of information about how the tiny, swirling eddies push and pull on the main flow. Without a good guess for this missing piece, the computer's prediction of speed, pressure, and temperature can drift far from reality. For decades, researchers have tried to fill this gap by teaching computers to learn the rules of turbulence from high-fidelity data, essentially showing the machine examples of how the swirling eddies behave. The hope is that if the computer learns these rules well, it can predict complex flows with much greater accuracy.

A team of researchers from the University of Stuttgart and the Gwangju Institute of Science and Technology has now refined how this learning process works. They discovered that simply showing the computer every example of turbulence with equal importance is not the most effective strategy. In their study, they found that different parts of the turbulent flow matter more than others, depending on what specific outcome the engineer wants to predict. To fix this, they developed a method that acts like a spotlight, shining brighter on the specific parts of the data that actually influence the final result, while dimming the parts that do not. By doing this, they were able to train a computer model that predicts complex flows, such as a jet of cooling air hitting a hot surface, with significantly better accuracy than previous methods, all without needing to run expensive, repeated simulations during the training phase.

The core of the problem lies in how computers currently learn from data. When training a model to understand turbulence, researchers typically feed it vast amounts of information about the swirling eddies, known as Reynolds stresses. The computer tries to minimize the error between its predictions and the real data. However, the standard approach treats every single piece of data as equally important. It assumes that a small error in one corner of the flow is just as bad as a small error in another. The researchers realized this is not true. In a complex flow, some parts of the turbulence have a huge impact on the final speed or temperature at a specific point, while other parts have almost no effect. If the computer spends too much time trying to perfectly match the unimportant parts, it may fail to get the important parts right. It is like trying to tune a radio by adjusting every single knob with the same intensity, rather than focusing on the ones that actually control the volume and clarity of the station you want to hear.

To solve this, the team introduced a technique called goal-oriented weighting. Before the computer even begins to learn, they run a single, specialized calculation that acts as a sensitivity test. This test asks a simple question: if I change the turbulence in this specific spot, how much does it change the speed of the air at the location I care about? The answer to this question generates a map of importance. Areas and types of turbulence that cause big changes in the final result get a high score, while those that cause little change get a low score. These scores are then used as weights during the training process. The computer is now instructed to pay extra attention to the high-scoring data and to be less concerned with the low-scoring data. This allows the model to focus its learning energy where it counts, effectively ignoring the noise that does not affect the final prediction.

The researchers tested this idea on three different types of flows to see if it worked. First, they looked at air moving through a square pipe, where the turbulence creates a secondary swirling motion that standard models often miss. By using their new weighting method, the computer successfully learned to predict this secondary motion much more accurately, matching the high-fidelity reference data closely. Next, they tested it on air flowing over a series of hills, a scenario where the flow separates and creates a wake of swirling air. Again, the weighted model outperformed the standard approach, predicting the size and shape of the separated flow region with greater precision. In both cases, the model learned to prioritize the specific components of turbulence that drive these phenomena, rather than trying to fit every detail equally.

The most challenging test came with a jet of cooling air injected into a crossflow, a situation common in turbine blades where hot gas must be kept cool. This flow is highly complex, with the jet separating, twisting, and mixing with the main stream in three dimensions. The researchers wanted to predict how well the cooling air would spread and protect the surface. They trained their model using data from a high-fidelity simulation at one specific operating condition and then tested it at a different condition to see if it could generalize. The standard method, which treated all data equally, improved the prediction of the jet's initial separation but failed to get the downstream flow and cooling distribution right. It predicted that the cooling air would spread in two separate, off-center patches, which was incorrect.

In contrast, the model trained with the goal-oriented weights got the downstream flow much closer to the real physics. It correctly predicted that the cooling air would stay more centered and spread more evenly, matching the high-fidelity reference data. Remarkably, this improvement extended to the cooling effectiveness itself, a measure of how well the surface is protected, even though the computer was never explicitly told to optimize for temperature. The model learned the correct momentum patterns, and the temperature followed naturally. When they tested this model on a different operating condition it had never seen before, it still outperformed the standard method, suggesting that the way it learned the flow physics was more robust.

The significance of this work lies in its efficiency and its precision. The researchers achieved these improvements without needing to run the expensive, time-consuming flow simulations repeatedly during the training process. They used a single, offline calculation to determine what was important, and then applied those lessons to the training. This means that engineers can now train better turbulence models faster, focusing on the data that actually matters for their specific goals. The study suggests that in the future, machine learning for physics should not just be about minimizing total error, but about minimizing the error that actually affects the decision at hand. By aligning the training process with the final prediction goal, the researchers have shown a path toward more reliable and efficient simulations for complex engineering challenges.

Technical Summary: Goal-Oriented Weighting of Reynolds-Stress Data for Learning Turbulence Models in Complex Flows

1. Problem Statement

Data-driven turbulence modeling for Reynolds-averaged Navier–Stokes (RANS) equations typically employs direct training (or a priori training), where constitutive relations for the Reynolds stress tensor are learned from high-fidelity data (DNS or LES) without solving the RANS equations during the optimization process. While computationally efficient compared to "solver-in-the-loop" approaches, standard direct training suffers from a critical misalignment: conventional loss functions apply uniform weights to all tensor components and spatial locations.

However, the influence of specific Reynolds-stress components and local errors on a downstream Quantity of Interest (QoI)—such as a specific velocity profile or aerodynamic force—varies significantly. In canonical shear flows, physical reasoning can identify relevant components (e.g., shear stress for mean velocity), but in complex 3D flows (e.g., jet-in-crossflow), the relevant stress components and spatial regions vary dynamically. Consequently, minimizing the aggregate reconstruction error of the Reynolds stress does not necessarily improve the prediction of the intended flow quantity. Conversely, indirect training methods that account for this influence via flow-level observations require repeated, expensive RANS solutions during optimization, creating a computational bottleneck.

2. Methodology

The authors propose a goal-oriented weighting framework that integrates QoI relevance into direct training while preserving solver-free parameter optimization. The core innovation is the use of a single offline adjoint sensitivity analysis to generate fixed weights for the supervised loss function.

Key Methodological Steps:

Adjoint Sensitivity Analysis: For a selected velocity-based QoI ( $J$ ), a single offline adjoint solution is computed based on the baseline RANS state. This quantifies the local linearized sensitivity of $J$ to perturbations in each Reynolds-stress anisotropy component ( $b_{ij}$ ) at every spatial location.
Weight Construction: The sensitivity magnitudes ( $G_{ij}$ $G_{ij}$ ) are converted into fixed weights ( $W_{ij}$ $W_{ij}$ ) for the training loss. The weight is decomposed into three factors:
1. Spatial Mask ( $M$ ): A binary mask that excludes regions with negligible sensitivity (e.g., freestream or plenum) to focus training on influential regions (e.g., injection holes, shear layers).
2. Cell Relevance ( $c$ ): A factor that assigns higher emphasis to cells with greater aggregate sensitivity.
3. Component Priority ( $w$ ): A factor that distributes emphasis among the six unique anisotropy components within a cell based on their relative sensitivities.
Training Loss: The modified loss function minimizes the weighted squared error between the predicted and target anisotropy:
$\mathcal{L}_{data} = \frac{1}{N_F N_c} \sum_{x \in \Omega_{map}} \sum_{(i,j) \in C} W_{ij}(x) [b_{ij}^{pred}(x) - b_{ij}^{target}(x)]^2$
Solver-Free Optimization: Once the weights are constructed, the model is trained using standard supervised learning. No further RANS or adjoint solutions are required during the parameter update loop.

3. Key Contributions

Goal-Oriented Direct Training: The paper introduces a method to align direct turbulence model training with downstream prediction goals without incurring the computational cost of iterative flow solutions.
Adjoint-Derived Weights: It demonstrates that a single offline adjoint evaluation is sufficient to identify which Reynolds-stress components and spatial locations are most critical for a specific QoI.
Spatially Varying Priorities: Unlike previous methods that might apply global component weights, this framework supports cellwise weights, allowing the model to prioritize different stress components in different flow regions (e.g., prioritizing $b_{xy}$ in shear layers vs. $b_{xz}$ inside a diffuser hole).
Handling Mismatch: For the jet-in-crossflow case, the authors utilize an "effective anisotropy target" that accounts for the mismatch between LES and RANS turbulent kinetic energy scales, enabling the learned closure to evolve naturally within the RANS transport equations.

4. Results

The framework was validated on three configurations: a square duct, flow over periodic hills, and a 7-7-7 diffuser jet-in-crossflow.

Square Duct (Secondary Motion):
- Goal: Predict in-plane velocity ( $U_y$ ).
- Finding: Adjoint weights correctly prioritized cross-plane normal stresses ( $b_{yy}, b_{zz}$ ) and shear stress ( $b_{yz}$ ).
- Outcome: The adjoint-weighted model reduced the reconstruction error of the normal-stress imbalance and provided more accurate secondary flow predictions compared to uniformly weighted training, with successful generalization to unseen Reynolds numbers.
Periodic Hill (Flow Separation):
- Goal: Predict streamwise velocity ( $U_x$ ) in the separated region.
- Finding: Weights heavily prioritized the shear anisotropy component $b_{xy}$ .
- Outcome: The adjoint-weighted model significantly reduced velocity errors in the separation region and predicted the reattachment point more accurately than the baseline and uniformly weighted models. This improvement generalized to different hill geometries ( $\alpha = 1.0, 1.2$ ).
Jet-in-Crossflow (Complex 3D Flow):
- Goal: Predict streamwise velocity ( $U_x$ ) and, indirectly, cooling effectiveness.
- Finding: The adjoint analysis revealed spatially varying priorities: $b_{xy}$ was critical in the downstream shear layer, while $b_{xz}$ was more relevant inside the diffuser hole.
- Outcome: While both learned models improved the internal hole separation, only the adjoint-weighted model corrected downstream velocity discrepancies (specifically near-wall deficits and shear layer overpredictions). Crucially, this velocity improvement led to a more accurate prediction of cooling effectiveness ( $\eta$ ) at both the training ( $BR=1.5$ ) and test ( $BR=1.0$ ) blowing ratios, despite the adjoint objective containing no thermal information.

5. Significance and Claims

The paper claims that reducing a uniformly aggregated constitutive error does not guarantee an improved a posteriori prediction. By weighting the training data according to their influence on the intended flow quantity, the framework provides a more targeted loss function.

The significance of this work lies in its ability to:

Bridge the gap between efficient direct training and flow-level performance goals.
Preserve computational efficiency by avoiding repeated RANS solves during optimization, relying instead on a single offline adjoint calculation.
Demonstrate transferability, showing that weights derived from one condition (e.g., specific Reynolds number or blowing ratio) can improve predictions in unseen conditions, suggesting that the identified "relevant" physics are robust.

The authors modestly note limitations: the weights are based on local linearized sensitivities and do not capture higher-order responses to large closure changes. Furthermore, improvements in thermal quantities (cooling effectiveness) are downstream responses of the coupled momentum-scalar system, not a direct optimization of the heat-flux model. Future work is suggested to explore iteratively updated sensitivities and coupled momentum-scalar adjoints.