Large-eddy simulation nets (LESnets) based on physics-informed neural operator for wall-bounded turbulence

This paper introduces LESnets, a physics-informed neural operator framework that integrates large-eddy simulation equations and wall models to enable accurate, efficient, and label-free long-term prediction of high-Reynolds-number wall-bounded turbulent flows on coarse grids.

The Gist

Imagine trying to predict how a chaotic river flows around a rock. The water swirls, eddies, and crashes in millions of tiny patterns. In the world of physics, this is called turbulence. When this happens next to a solid wall (like a pipe or an airplane wing), it's called wall-bounded turbulence.

Predicting this is incredibly hard. Traditional computer simulations are like trying to count every single water molecule; they are accurate but take so much computing power that they are often impossible for big, real-world problems.

This paper introduces a new "smart" way to solve this problem using Artificial Intelligence (AI). Here is the breakdown of what they did, using simple analogies:

1. The Problem: The "Too Big to Count" Puzzle

Imagine trying to predict the weather for next month. You have a map, but the details are so tiny (like individual raindrops) that your computer crashes if it tries to track them all.

• Traditional AI (Data-Driven): Usually, AI learns by looking at millions of "answer keys" (labeled data). It memorizes patterns. But in fluid dynamics, getting those "answer keys" is as expensive and slow as running the super-hard simulations we are trying to avoid.
• The Challenge: Wall-bounded turbulence is messy. The water behaves very differently right next to the wall compared to the middle of the stream. Standard AI models often get confused here and make mistakes over time.

2. The Solution: "LESnets" (The Physics-Savvy Student)

The authors created a new AI model called LESnets. Think of it not as a student who just memorizes flashcards, but as a student who has the textbook (Physics) open in front of them while they study.

• No Answer Keys Needed: Unlike most AI that needs a huge library of pre-solved examples to learn, LESnets learns by trying to satisfy the laws of physics (like conservation of mass and momentum). It's like a student who solves a math problem by checking if the answer makes sense according to the rules of algebra, rather than copying from a cheat sheet.
• The "Hard Constraint" Rule: Imagine a train track. The train must stay on the tracks. In this AI, the "walls" of the pipe are like the tracks. The model is built so that it is physically impossible for the water to go through the wall. This is called a "hard constraint," and it stops the AI from making silly mistakes near the edges.

3. The Secret Sauce: The "Wall Model"

When water flows fast next to a wall, it creates a very thin, chaotic layer that is hard to see on a coarse (low-resolution) map.

• The Analogy: Imagine trying to see the texture of a brick wall from a helicopter. You can't see the individual bricks.
• The Fix: The authors added a "Wall Model." This is like a rulebook that tells the AI: "Even though you can't see the tiny bricks from this height, you know the wall is rough, so the water should slow down right next to it." This allows the AI to use a low-resolution map (which is fast) but still get the physics near the wall right.

4. How It Works: The "Self-Correcting" Loop

The AI doesn't just guess once. It works like a video game character that moves one step, checks the rules, and then moves again.

1. Predict: It guesses what the water flow will look like a split-second later.
2. Check: It checks its guess against the laws of physics (the textbook).
3. Adjust: If the guess breaks the laws of physics, it learns from that mistake and updates its "brain" (the neural network).
4. Repeat: It does this over and over to predict the flow for a long time.

5. The Results: Fast and Accurate

The researchers tested this on "turbulent channel flows" (water flowing in a pipe) at three different speeds (Reynolds numbers).

• Speed: The AI model was much faster than traditional high-precision simulations. It could predict the flow in seconds that would take a supercomputer hours to calculate.
• Accuracy: Even though it was fast, it was just as accurate as the traditional methods for predicting how the water moves, the speed of the flow, and the swirling patterns (vortices).
• Bonus Feature: The model can even "learn" the right settings for its own physics rules automatically. If it doesn't know a specific coefficient (a number that defines how the fluid behaves), it can figure it out while it trains, using just a tiny bit of extra data.

Summary

The paper presents LESnets, a new type of AI that predicts how turbulent fluids flow near walls. Instead of needing a massive library of pre-solved examples, it learns by strictly following the laws of physics. It uses a special "rulebook" for the walls to stay accurate even when using low-resolution maps. The result is a tool that is fast, accurate, and doesn't need expensive training data, making it a powerful new way to simulate complex fluid flows like those in pipes or around aircraft.

Technical Summary

1. Problem Statement

Predicting three-dimensional (3D) wall-bounded turbulent flows is a significant challenge for machine learning (ML) methods, particularly in scenarios with limited labeled data and high Reynolds numbers ($Re_\tau$).

• Limitations of Traditional CFD: Direct Numerical Simulation (DNS) is computationally prohibitive at high $Re$. Reynolds-Averaged Navier-Stokes (RANS) and standard Large-Eddy Simulation (LES) reduce costs but still struggle with extensive parametric design spaces and near-wall resolution requirements.
• Limitations of Existing ML:
  • Data-driven Neural Operators (e.g., FNO, DeepONet): Require massive amounts of high-resolution labeled data, which is expensive to generate. They often lack physical consistency and struggle with long-term stability in chaotic turbulent flows.
  • Physics-Informed Neural Networks (PINNs): While they enforce physical laws, they often fail to train stably for multi-scale turbulent flows due to strong nonlinearity and sensitivity to initial conditions.
  • Wall-Bounded Turbulence Specifics: Existing Physics-Informed Neural Operator (PINO) methods struggle with the anisotropic nature of wall-bounded flows, non-periodic boundary conditions, and the "log-layer mismatch" problem when using coarse grids.

2. Methodology: LESnets Framework

The authors propose LESnets, a novel framework that integrates Large-Eddy Simulation (LES) equations into a Factorized Fourier Neural Operator (F-FNO). The key methodological innovations include:

A. Architecture and Operator Learning

• Base Model: Uses F-FNO instead of standard FNO. F-FNO decomposes the 3D Fourier transform into three 1D transforms along each spatial dimension ($x, y, z$). This drastically reduces the number of parameters from $O(k_1 k_2 k_3)$ to $O(k_1 + k_2 + k_3)$, making it efficient for 3D turbulence.
• Encoder-Decoder: Replaces standard projection layers with convolutional encoder-decoder blocks to handle grid information implicitly, reducing model size.
• Hard Constraints: Unlike traditional PINNs that use "soft" constraints via loss functions, LESnets enforces no-slip boundary conditions as hard constraints. The velocity at the walls is explicitly set to zero during the inference and training steps, ensuring physical consistency at boundaries without relying on penalty weights.

B. Physics-Informed Loss Function

The model is trained in a fully self-supervised manner (no labeled flow field data required for the primary training).

• Governing Equations: The loss function is derived from the filtered incompressible Navier-Stokes equations (LES equations) using a Finite Difference (FD) scheme in physical space (rather than automatic differentiation, which is memory-intensive).
• Subgrid Scale (SGS) Modeling: Incorporates the Wall-Adapting Local Eddy-Viscosity (WALE) model.
• Wall Model Integration: For high Reynolds numbers, a wall model (based on the law of the wall) is integrated into the loss function. This approximates the log-law profile, allowing the model to use coarse grids near the wall without resolving the viscous sublayer, significantly reducing computational cost.

C. Training and Inference Strategies

• Training: The model minimizes the PDE residual loss ($L_{PDE}$). If a small amount of filtered DNS (fDNS) data is available, the WALE coefficient ($C_w$) can be treated as a learnable parameter and optimized simultaneously.
• Inference: Uses an autoregressive strategy. The predicted velocity field at time $t$ is fed back as the initial condition for time $t+1$, enabling long-term temporal evolution prediction.

3. Key Contributions

1. First PINO for 3D Wall-Bounded Turbulence: This is the first application of a Physics-Informed Neural Operator framework specifically designed for 3D wall-bounded turbulent flows at high Reynolds numbers without labeled data.
2. Data-Free Training: The framework eliminates the need for large labeled datasets by relying on the physics of the LES equations and hard boundary constraints.
3. Coarse-Grid Efficiency: By integrating a wall model, LESnets achieves accurate predictions at high $Re$ (up to $Re_\tau \approx 1000$) using coarse grids, bypassing the computational cost of Wall-Resolved LES (WRLES).
4. Automatic SGS Coefficient Optimization: The method allows for the automatic learning of the SGS model coefficient ($C_w$) using a small subset of fDNS data, reducing reliance on empirical tuning.
5. Flexible Temporal Output: The model can generate time-series outputs at flexible time intervals without retraining, a feature difficult for standard data-driven models.

4. Results and Performance

The model was evaluated on turbulent channel flows at three friction Reynolds numbers: $Re_\tau \approx 180, 590,$ and $1000$. It was compared against:

• Baselines: Traditional LES (WALE), Implicit U-Net enhanced FNO (IUFNO), and standard F-FNO.
• Metrics: Mean velocity profiles, Reynolds stresses, RMS fluctuations, kinetic energy spectra, and instantaneous vortex structures (Q-criterion).

Key Findings:

• Accuracy: LESnets outperformed IUFNO and F-FNO in predicting long-term statistics. At $Re_\tau \approx 180$ and $590$, LESnets results were nearly identical to traditional WALE-LES. At $Re_\tau \approx 1000$, LESnets with the wall model (LESnets-WM) significantly outperformed data-driven baselines, correctly capturing the log-law region where others failed.
• Stability: The autoregressive inference remained stable over 395 time steps, whereas IUFNO showed deviations and non-physical behavior in energy spectra.
• Efficiency: LESnets achieved computational speeds comparable to F-FNO (approx. 4–15 seconds for 10,000 time steps) but with significantly higher accuracy than data-driven models and much lower cost than traditional LES (which took 38–140 seconds on CPU cores).
• SGS Coefficient Learning: When $C_w$ was unknown, the model successfully learned a value ($C_w \approx 0.225$) that produced results consistent with DNS and WALE, demonstrating the capability for inverse problem solving.

5. Significance

This work represents a major step forward in scientific machine learning (SciML) for fluid dynamics:

• Bridging the Gap: It successfully bridges the gap between data-driven operator learning and physics-based turbulence modeling, offering a solution that is both data-efficient and physically consistent.
• Scalability: The ability to simulate high-Reynolds-number wall-bounded flows on coarse grids using a neural operator suggests a pathway toward real-time or near-real-time turbulence prediction for engineering applications (e.g., aerospace design, weather modeling).
• Generalizability: The framework's ability to handle non-periodic boundaries and optimize model coefficients automatically makes it a robust candidate for complex flow problems where labeled data is scarce or non-existent.

In conclusion, LESnets demonstrates that integrating physical constraints (LES equations, wall models) directly into the architecture and loss function of neural operators can overcome the stability and accuracy limitations of current ML methods for complex, multi-scale turbulent flows.