A hybrid framework integrating physics-informed neural operators with conventional methods for turbulence
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "A hybrid framework integrating physics-informed neural operators with conventional methods for turbulence".
Mira: Here's the gist of the paper. Large-eddy simulation nets (LESnets) based on physics-informed neural operator for wall-bounded turbulence models with hard constraints and automatic subgrid scale model optimization to enable accurate long-term predictions without labeled data Model Development and Framework The authors developed the LESnets framework which integrates large-eddy simulation (LES) equations into the factorized Fourier neural operator (F-FNO) for wall-bounded turbulence The LESnets framework does not rely on labeled data for training, which enables it to generate temporal solutions over flexible time horizons during the training process By imposing boundary conditions as hard constraints, they improve the performance of the model near boundary The framework incorporates a wall model into the loss function for the physics-informed loss, thus enabling reliable simulations of wall-bounded turbulence at high Reynolds number using coarse grids.
Kai: Training Signal Construction The training data is structured by sampling LES of 3D turbulent channel flows at three friction Reynolds numbers: Reτ ≈ 180, 590, and 1000 over a duration time of 200∆T.
Mira: The training datasets A180 train, A590 train, and A1000 train are constructed from different initial conditions to form the training datasets.
Kai: For each dataset, 20 groups of WALE or WMLES instantaneous velocity and pressure flow fields are generated using different initial conditions to construct the training datasets.
Paper summary: Kai: So this paper, "A hybrid framework integrating physics-informed neural operators with conventional methods for turbulence," it’s about using these large-eddy simulation nets, or LESnets, to model wall-bounded turbulence. The main thing they claim is that you can do this without needing labeled data.
Mira: Right, so the thesis here seems to be building a framework that combines physics-informed neural operators with conventional methods for these turbulent flows. They specifically focus on wall-bounded turbulence and want to show how you can get accurate long-term predictions even when data is scarce.
Kai: Exactly, and they tackle the problem of multi-scale vortex structures in wall-bounded turbulence, which makes it tough for most physics constrained neural operator methods to make stable and accurate long-term predictions at high Reynolds numbers <ref:2604.26621#pg2>.
Lev: But on the practical side, they are using large-eddy simulation equations and integrating them into a factorized Fourier neural operator, or F-FNO. That sounds computationally heavy for real hardware to run efficiently.
Mira: Well, what’s interesting is that they use hard constraints as boundary conditions. This helps improve the model's performance right near the wall <ref:2604.26621#pg1>. And they incorporate a wall model directly into their physics-informed loss function, which lets them simulate high Reynolds number flows on coarse grids.
Kai: So, it’s trying to get that accuracy by having the physics guide the learning process instead of just relying on data to teach the network everything.
Lev: But I wonder about how they handle that subgrid-scale model optimization during training. They mentioned optimizing the coefficient of the subgrid-scale model using only a single set of flow data <ref:2604.26621#pg1>.
Mira: That’s where my concern lies for me, Lev. If you're only using one set of flow data to optimize that model coefficient, are you sure it generalizes well across different flow regimes? The assumption there is pretty strong.
Kai: They treat that subgrid-scale model coefficient as a learnable parameter in the neural network by incorporating fDNS data and adding an extra term to the loss function when it's unknown <ref:2604.26621#pg1>. That’s their way of letting the AI learn that part.
Lev: From an error correction point, if that coefficient isn't perfectly learned, how much overhead does that add when you try to run this on real hardware? We need to know where the computational cost lands when you have these physics constraints layered on top of a neural operator.
Mira: The paper claims the performance is comparable to other data-driven models like IUFNO and F-FNO in terms of prediction accuracy and efficiency <ref:2604.26621#pg1>. But they also say it achieves prediction accuracy comparable to traditional LES methods while offering higher computational efficiency <ref:2604.26621#pg1>.
Kai: They also highlight that the LESnets-WM model outperforms its data-driven counterpart in long-term predictions of turbulent channel flow at high Reynolds numbers <ref:2604.26621#pg1>.
Conclusion: Kai: So to wrap up, this paper, "A hybrid framework integrating physics-informed neural operators with conventional methods for turbulence," it’s essentially proposing a way to use physics constraints within a neural operator structure to model wall-bounded turbulence without needing massive amounts of labeled data.
Mira: The core implication is that you can get reliable long-term predictions in these complex turbulent flows, even at high Reynolds numbers, using coarse grids where direct simulation or full direct numerical simulation is impractical <ref:2604.26621#pg3>.
Lev: What this means for running this on real hardware is that the overhead from the physics-informed constraints and the subgrid-scale optimization needs to be managed very carefully if you want to achieve those claimed computational efficiencies.
Kai: And the authors are focused on showing that they can output results at a flexible time interval, meaning you can get predictions for whatever time horizon you actually need, as long as it’s not too far out <ref:2604.26621#pg1>.
Mira: The integration of the wall model into the loss function is key here. It’s not just a mathematical trick; it's an attempt to enforce the law of the wall physically within the model, which should help stabilize those tricky near-wall regions <ref:2604.26621#pg1>.
Kai: Overall, this work shows that combining physics-informed neural operators with conventional methods offers a path toward more efficient and accurate turbulence modeling for real engineering problems where data is limited <ref:2604.26621#pg3>.
Sunan Zhao, Yunpeng Wang, Huiyu Yang, Zhihong Guo, Jianchun Wang
Southern University of Science and Technology, Shenzhen
physics.flu-dyn
Submitted: 2026-04-29
Updated: 2026-10-29
Code: https://github.com/Turbulence-AI/IUFNOCHL
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 79/100
The gist: Large-eddy simulation nets (LESnets) based on physics-informed neural operator for wall-bounded turbulence models with hard constraints and automatic subgrid scale model optimization to enable…
Terminology
Summary
Large-eddy simulation nets (LESnets) based on physics-informed neural operator for wall-bounded turbulence models with hard constraints and automatic subgrid scale model optimization to enable accurate long-term predictions without labeled data
Model Development and Framework
The authors developed the LESnets framework which integrates large-eddy simulation (LES) equations into the factorized Fourier neural operator (F-FNO) for wall-bounded turbulence The LESnets framework does not rely on labeled data for training, which enables it to generate temporal solutions over flexible time horizons during the training process By imposing boundary conditions as hard constraints, they improve the performance of the model near boundary The framework incorporates a wall model into the loss function for the physics-informed loss, thus enabling reliable simulations of wall-bounded turbulence at high Reynolds number using coarse grids.
Training Signal Construction
The training data is structured by sampling LES of 3D turbulent channel flows at three friction Reynolds numbers: Reτ ≈ 180, 590, and 1000 over a duration time of 200∆T. The training datasets A180 train, A590 train, and A1000 train are constructed from different initial conditions to form the training datasets. p. 6 For each dataset, 20 groups of WALE or WMLES instantaneous velocity and pressure flow fields are generated using different initial conditions to construct the training datasets. p. 6
Optimization Strategy
The LESnets framework optimizes the coefficient of the subgrid-scale (SGS) model during the training process by incorporating only a single set of flow data. The physics-informed loss function is defined as L(θ) = LPDE(θ) + λ · LS GS (θ), where the SGS loss is FD[R(u(i)f DNS;Cw(θ))]. p. 6 When the coefficient of the WALE model Cw is unknown, it can be regarded as a learnable parameter in neural networks by incorporating fDNS data and adding an additional term in the loss function. p. 9
Performance and Validation
The LESnets model's performance is comparable to two data-driven models, namely the implicit U-Net enhanced Fourier neural operator (IUFNO) and F-FNO in terms of prediction accuracy and efficiency. Meanwhile, the LESnets model achieves prediction accuracy comparable to traditional LES methods while offering a higher computational efficiency. p. 1 The LESnets-WM model outperforms its data-driven counterpart in long-term predictions of turbulent channel flow at high Reynolds numbers. p. 3
Flexibility and Efficiency
The LESnets framework allows the model to output results at a flexible time interval. The authors demonstrated that regardless of the chosen time interval, LESnets remain in good agreement with the WALE results over short time horizons. p. 1 The computational cost of LESnets is much faster than the WALE methods, similar to previous studies. The framework preserves the key advantage of not relying on any labeled data for supervision. p. 1
Wall Model Integration
The authors incorporate a wall model into the LESnets to approximately enforce the law of the wall as a physical constraint within the LESnets model. This involves approximating the log-law in a segmented manner using Eq. p. 4 (11) and employing the sixth grid point away from the wall in the wall-normal direction as a sampling location for the wall model. p. 5 The LESnets-WM model is able to outperform its data-driven counterpart in long-term predictions of turbulent channel flow at high Reynolds numbers. p. 3
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Automated SGS model coefficient learning: The LESnets framework can automatically optimize the coefficient of the subgrid-scale model during training by incorporating a subset of fDNS data, as stated in
Employing a single set of flow data as a priori, the coefficient of the subgrid scale model can be automatically optimized during the training process for wall-bounded turbulence.
This allows systems to achieve more accurate predictions without needing prior knowledge or manual tuning. -
Physics-constrained boundary enforcement: The LESnets framework improves performance near boundaries by
imposing boundary conditions as hard constraints,
which facilitates training and can potentially improve generalization, as mentioned in Section 3.3.3 of the paper. This means the AI system can be trained to strictly adhere to physical wall conditions like no-slip without relying solely on soft loss functions. -
High Reynolds number prediction: The LESnets model enables
reliable simulations of wall-bounded turbulence at high Reynolds number using coarse grids
by integrating a wall model for the physics-informed loss, which addresses the challenge thatthe LES equations fail to accurately capture near-wall flow characteristics on coarse grids without the incorporation of appropriate wall models for high-Reynolds-number wall-bounded turbulence.
-
Flexible temporal output: The LESnets framework allows the model to
output results at a flexible time interval
during inference, which is achieved through an autoregressive strategy, enabling long-term prediction byiteratively feeding the predicted velocity field back into the model as a new initial field.
This capability extends prediction horizons beyond fixed time steps. -
Enhanced architecture stability: The system can utilize the Factorized Fourier neural operator (F-FNO) within LESnets to
ensure the stability and accuracy of long-term predictions for wall-bounded turbulent flows,
addressing limitations whereemploying a simple FNO architecture as the neural operator can lead to degraded performance in wall-bounded turbulent flows.
Sources
- Fourier Neural Operator for Parametric Partial Differential Equations
- Simulating Three-dimensional Turbulence with Physics-informed Neural Networks
- A generative model for age and income distribution
- Neural Operator: Learning Maps Between Function Spaces
- Adaptive Fourier Neural Operators: Efficient Token Mixers for Transformers
- Factorized Fourier Neural Operators
- Strategies for Pretraining Neural Operators
- Learning Physical Operators using Neural Operators
- OmniFluids: Physics Pre-trained Modeling of Fluid Dynamics
- Solving forward and inverse PDE problems on unknown manifolds via physics-informed neural operators
- Physics-enhanced Neural Operator for Simulating Turbulent Transport
- A Physics-informed Multi-resolution Neural Operator
- Physics-Informed Transformer operator for the prediction of three-dimensional turbulence
- PEST: Physics-Enhanced Swin Transformer for 3D Turbulence Simulation
- Automatic differentiation in machine learning: a survey
- Fourier Neural Operator with Learned Deformations for PDEs on General Geometries
- SOAP: Improving and Stabilizing Shampoo using Adam
Related papers
- Data-driven impeller model for efficient large eddy simulations of metastable von K'arm'an flows
- Numerical study of nitrogen oxides chemistry during plasma assisted combustion in a sequential combustor
- Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence
- Early onset of secondary shear instability in Kelvin-Helmholtz braids at high Reynolds number
- Rotational covariance, continuous spectrum and non-modal growth of three-dimensional planar weak viscous shock layers
- Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows