Machine-learning-based simulation of turbulent flows over periodic hills using a hybrid U-Net and Fourier neural operator framework

This paper proposes a hybrid U-Net and Fourier neural operator (HUFNO) framework that effectively combines convolutional and Fourier-based approaches to achieve highly accurate and computationally efficient large-eddy simulations of turbulent flows over periodic hills, outperforming traditional models and demonstrating strong generalization across unseen conditions.

The Gist

Imagine you are trying to predict how a river flows over a series of smooth, rolling hills. In the real world, water doesn't just slide smoothly; it swirls, crashes, and creates chaotic whirlpools (turbulence). To understand this for engineering—like designing better airplanes or predicting weather—we need to simulate these flows on a computer.

However, doing this with traditional math is like trying to count every single grain of sand on a beach to predict how the tide moves. It's incredibly accurate, but it takes so much computer power that it's often impossible to do quickly.

This paper introduces a new, super-smart "AI weather forecaster" that solves this problem. Here is the story of how they did it, explained simply.

The Problem: The "One-Size-Fits-All" Mistake

For years, scientists have tried to use Artificial Intelligence (AI) to predict these flows. They tried two main approaches:

1. The "Pattern Matcher" (U-Net): This is like a student who is great at memorizing a specific map. It's very good at looking at a local neighborhood and saying, "I know what happens right here." But if you move the map or change the shape of the hills, it gets confused.
2. The "Global Predictor" (Fourier Neural Operator or FNO): This is like a student who understands the laws of physics globally. It can predict how waves move across the whole ocean. But it has a flaw: it assumes the world is a perfect circle where the left side connects seamlessly to the right side. In real life, hills have ends and starts; they aren't perfect circles. When you force this model to look at a hill, it gets "dizzy" because the math doesn't fit the edges.

The Solution: The "Hybrid Chef" (HUFNO)

The authors created a new model called HUFNO (Hybrid U-Net and Fourier Neural Operator). Think of this model as a master chef who hires two specialized sous-chefs to run the kitchen:

• Chef FNO (The Global Specialist): This chef handles the parts of the river that flow in a perfect loop (like the river flowing left-to-right over the hills). Because the river repeats itself in that direction, Chef FNO uses its "global laws" to predict the flow instantly and efficiently.
• Chef U-Net (The Local Specialist): This chef handles the tricky parts where the river hits the solid ground, the top of the hill, or the bottom of the valley. These areas don't repeat; they are unique and messy. Chef U-Net zooms in, looks at the specific details of the hill's shape, and figures out exactly how the water swirls there.

By combining them, the model gets the best of both worlds: the speed of the global predictor and the accuracy of the local observer.

The Test: The "Periodic Hill" Challenge

To test this new AI, the scientists used a classic problem: Turbulent flow over periodic hills.
Imagine a long, endless row of identical hills. The wind blows over them, creating massive separation zones where the air detaches from the hill and swirls back. This is a nightmare for traditional computers because the air behaves chaotically.

They trained their AI on data from a super-accurate (but slow) simulation called DNS (Direct Numerical Simulation). Then, they let the AI run the simulation on its own.

The Results: Speed and Smarts

The results were impressive. The HUFNO model didn't just guess; it learned the physics.

1. It's a Speed Demon: Traditional methods took hours of supercomputer time to simulate a few seconds of wind. The HUFNO model did the same job in seconds. It was roughly 100 times faster than the traditional methods, even though it was running on a single graphics card.
2. It's More Accurate: When compared to the "gold standard" (the slow, perfect simulation), the HUFNO model predicted the speed of the wind, the pressure, and the swirling eddies much better than the old-school math models (like Smagorinsky or WALE).
3. It's a Chameleon (Generalization): This is the coolest part. The AI was trained on specific hill shapes and wind speeds. Then, the scientists tricked it:
   • New Hills: They showed it hills it had never seen before (steeper or flatter). The AI still worked perfectly.
   • New Speeds: They changed the wind speed to a level it hadn't seen. The AI adapted.
   • New Shapes: They even tested it on 3D hills (like a dune) instead of just 2D ridges. It handled the complexity with ease.

Why This Matters

Think of this technology as a flight simulator for engineers.

• Before: If you wanted to design a car or a plane, you had to build a physical model, put it in a wind tunnel, or run a simulation that took days to finish.
• Now: With HUFNO, you could change the shape of the car, the speed of the wind, or the terrain, and get a highly accurate prediction of how the air will behave in a fraction of a second.

The paper concludes that this "Hybrid Chef" approach is a huge leap forward. It allows us to simulate complex, swirling flows over curved surfaces (like mountains, dams, or airplane wings) with a speed and accuracy that was previously impossible. It's not just a faster calculator; it's a new way of teaching computers to "feel" the flow of fluids.

Technical Summary

1. Problem Statement

Simulating massively separated turbulent flows over curved bodies is a significant challenge in computational fluid dynamics (CFD), particularly for Large-Eddy Simulation (LES). Traditional LES relies on subgrid-scale (SGS) models (e.g., Smagorinsky or WALE) to approximate unresolved scales, which often struggle with accuracy in complex flow separations and curved boundaries.

While Machine Learning (ML) offers a potential alternative, existing neural network (NN) approaches face limitations:

• Generalization: Standard NNs often fail to generalize across different boundary conditions or Reynolds numbers.
• Periodicity Assumption: Fourier Neural Operators (FNO) are highly efficient for periodic domains but struggle with non-periodic boundaries (common in curved geometries) unless treated with ad-hoc methods.
• Complexity: 3D turbulent flows involve complex nonlinear interactions, requiring high model capacity and memory, which can lead to instability or overfitting.

The authors aim to develop a surrogate model for LES that can rapidly and accurately predict strongly separated 3D turbulent flows over curved boundaries (specifically periodic hills) while maintaining transferability to unseen initial conditions, Reynolds numbers, and hill geometries.

2. Methodology: The HUFNO Framework

The authors propose a Hybrid U-Net and Fourier Neural Operator (HUFNO) framework. This architecture strategically combines the strengths of Convolutional Neural Networks (CNN) and FNO to handle mixed boundary conditions.

• Core Concept: The flow field is decomposed based on its periodicity:
  • Periodic Directions (Streamwise $x$ and Spanwise $z$): Handled by the Fourier Neural Operator (FNO). FNO learns mappings in the frequency domain, offering high efficiency and accuracy for periodic data.
  • Non-Periodic Direction (Wall-normal $y$): Handled by a U-Net (CNN-based). The U-Net processes the non-periodic direction, utilizing its encoder-decoder structure and skip connections to capture local features and boundary effects without assuming periodicity.
• Architecture Details:
  • The model takes the velocity field and hill shape as inputs.
  • It predicts the velocity increment ($v(m+1) - v(m)$) over a time step $\Delta T$.
  • The iterative update rule (Eq. 7) integrates the Fourier operations in $x$ and $z$ with the U-Net operation in $y$, incorporating residual connections to stabilize gradient flow.
  • Unlike previous Implicit U-Net enhanced FNO (IUFNO) models where FNO was applied in all directions, HUFNO strictly isolates FNO to periodic directions, reducing parameter count and improving robustness for wall-bounded flows.
• Training Data:
  • Generated using Direct Numerical Simulation (DNS) of turbulent flow over periodic hills at Reynolds numbers $Re = 700, 1400, \text{and } 5600$.
  • DNS data was coarsened to LES grids to serve as the ground truth for training the surrogate model.
  • The training set included variations in initial conditions and hill shapes (controlled by a shape parameter $a$).

3. Key Contributions

1. Novel Hybrid Architecture: Introduction of HUFNO, which uniquely partitions the domain processing: FNO for periodic directions and U-Net for non-periodic (wall-normal) directions. This addresses the specific challenge of applying spectral methods to wall-bounded flows with curved boundaries.
2. Surrogate LES Model: Development of a data-driven LES surrogate that outperforms traditional SGS models (Smagorinsky and WALE) in predicting turbulence statistics, energy spectra, and flow separation structures.
3. Demonstrated Generalization: The model was rigorously tested on unseen scenarios, proving its ability to generalize to:
   • Unseen initial conditions.
   • Unseen hill slopes (geometries).
   • Unseen Reynolds numbers (trained on $Re=700, 5600$, tested on $Re=1400$).
   • Fully 3D hill geometries (where shape varies in both streamwise and spanwise directions).
4. Computational Efficiency: The HUFNO framework achieves a massive reduction in computational cost compared to traditional CPU-based LES, running on a single GPU.

4. Results

The performance of HUFNO was evaluated against DNS benchmarks, traditional LES (Smagorinsky/SMAG and WALE models), and other NN baselines (FNO, U-Net, IUFNO).

• Accuracy:
  • Turbulence Statistics: HUFNO showed the closest agreement with DNS for mean velocity, Reynolds stresses ($\langle u'u' \rangle, \langle u'v' \rangle$), and wall shear stress. Traditional LES models (especially SMAG) showed significant deviations.
  • Energy Spectrum: HUFNO accurately recovered the multiscale energy distribution, whereas traditional models tended to overestimate large-scale energy and underestimate small-scale energy.
  • Flow Topology: Analysis of the $Q-R$ invariant map (characterizing flow topology) showed HUFNO predictions closely matched DNS, while SMAG significantly underestimated extreme values.
  • Separation Structures: HUFNO accurately captured flow separation bubbles and re-circulation streamlines, outperforming all other models.
• Generalization:
  • The model maintained high accuracy when tested on hill shapes and Reynolds numbers not present in the training set.
  • While generalization to unseen Reynolds numbers was slightly more challenging than unseen initial conditions, HUFNO still outperformed traditional LES.
• Efficiency:
  • Speed: HUFNO simulations were orders of magnitude faster than traditional LES. For example, at $Re=5600$, HUFNO took 6.33 seconds per 10,000 time steps on a GPU, compared to 270–391 seconds for traditional LES on 64 CPU cores.
  • Scalability: The efficiency advantage is even more pronounced when considering that traditional LES requires significant CPU resources, while HUFNO runs on a single GPU.

5. Significance and Future Outlook

• Impact: This work represents a significant step toward replacing or augmenting traditional LES for complex, separated flows. It demonstrates that ML-based surrogates can handle the complexity of 3D wall-bounded turbulence with curved boundaries, a scenario where previous NN methods struggled.
• Applications: The framework is highly relevant for engineering applications involving bluff bodies, mountainous terrain, dams, and urban microclimates, where non-periodic boundary conditions and flow separation are prevalent.
• Limitations & Future Work:
  • The current model assumes Cartesian grids; extension to irregular/non-Cartesian grids (common in real engineering) is needed.
  • Performance may degrade with extremely steep geometries or insufficient training data at very high Reynolds numbers.
  • Future directions include integrating physics-informed constraints (PINNs), geometry-informed neural operators (GINO), and data assimilation techniques to further enhance robustness and reduce data requirements.

In conclusion, the HUFNO framework successfully bridges the gap between the efficiency of spectral methods and the flexibility of CNNs, offering a powerful, fast, and accurate tool for simulating complex turbulent flows.