Source-linked AI summary
Towards Physics-informed Deep Learning for Turbulent Flow Prediction
Rui Wang, Karthik Kashinath, Mustafa Mustafa, Adrian Albert, Rose Yu
TL;DR
The paper tackles accurate prediction of chaotic turbulent flows while addressing the limited physical knowledge in conventional deep learning. It combines RANS–LES turbulence modeling with trainable spectral filters and specialized neural networks, achieving lower error and improved physical consistency up to 60 frames ahead.
Problem
Deep learning methods generally incorporate little physical knowledge, while turbulent flow is chaotic, multiscale, and high-dimensional, making accurate prediction difficult.
Method
TF-Net combines RANS–LES coupling with trainable spectral filters and specialized U-nets to model turbulent-flow components at different scales.
Results
TF-Net achieves significant improvements over state-of-the-art baselines, including 11.1% lower prediction RMSE and better energy-spectrum, turbulence-kinetic-energy, and divergence measures.
Takeaways & Limitations
The model produces accurate turbulent-flow forecasts that preserve physically meaningful quantities, including divergence, turbulence kinetic energy, and energy spectrum.
Takeaways & Limitations
The reported setting uses a Boussinesq approximation for Rayleigh–Bénard convection, producing a divergence-free flow.
Abstract
from arXiv · showhide
While deep learning has shown tremendous success in a wide range of domains, it remains a grand challenge to incorporate physical principles in a systematic manner to the design, training, and inference of such models. In this paper, we aim to predict turbulent flow by learning its highly nonlinear dynamics from spatiotemporal velocity fields of large-scale fluid flow simulations of relevance to turbulence modeling and climate modeling. We adopt a hybrid approach by marrying two well-established turbulent flow simulation techniques with deep learning. Specifically, we introduce trainable spectral filters in a coupled model of Reynolds-averaged Navier-Stokes (RANS) and Large Eddy Simulation (LES), followed by a specialized U-net for prediction. Our approach, which we call turbulent-Flow Net (TF-Net), is grounded in a principled physics model, yet offers the flexibility of learned representations. We compare our model, TF-Net, with state-of-the-art baselines and observe significant reductions in error for predictions 60 frames ahead. Most importantly, our method predicts physical fields that obey desirable physical characteristics, such as conservation of mass, whilst faithfully emulating the turbulent kinetic energy field and spectrum, which are critical for accurate prediction of turbulent flows.
1 INTRODUCTION
The paper addresses turbulent-flow prediction by combining physics-informed modeling with deep learning for a highly chaotic, multiscale, high-dimensional system. It introduces TF-Net, a hybrid RANS–LES and neural-network framework that improves forecasting accuracy and preserves important physical quantities.
- Motivation: Existing deep learning methods are primarily statistical and generally incorporate little underlying physical knowledge.Prior physics-informed approaches often use ad-hoc regularization or focus only on spatial modeling.
- Motivation: Turbulent flow prediction is difficult because the dynamics are chaotic, multiscale, high-dimensional, and sensitive to initial conditions.A discretized 128 × 128 × 100 velocity field has 10^6 dimensions and is prone to error propagation.
- Method: TF-Net unifies RANS–LES coupling with custom deep neural networks in a hybrid framework for turbulent-flow forecasting.The framework decomposes turbulent flow into scale components and uses specialized U-nets for prediction.
- Results: TF-Net forecasts velocity fields up to 60 steps ahead while preserving accurate and physically meaningful predictions.The paper evaluates physical quantities including divergence, turbulence kinetic energy, and energy spectrum.
- Results: 11.1% reduction in prediction RMSE and 30.1% improvement in the energy spectrum were achieved compared with the best baseline.The reported comparison also includes 21% turbulence kinetic energy RMSEs and 64.2% reduction of flow divergence in difference from the target.
2 RELATED WORK
Related work spans physics-based and data-driven spatiotemporal forecasting, turbulence modeling, fluid animation, and video prediction. The paper distinguishes TF-Net by targeting real-time turbulent-flow evolution while incorporating physical structure.
- Spatiotemporal Forecasting: Spatiotemporal forecasting methods range from differential-equation models to deep learning, with recent work adding physical constraints through loss regularization.Such regularization is described as ad hoc and difficult to tune.
- Turbulence Modeling: Prior turbulence-modeling studies learn Reynolds stresses, generate turbulent flows, or accelerate simulation, but do not provide the same forecasting setting as TF-Net.One cited generative model cannot make real-time future predictions from historical data, while another hybrid framework does not model turbulence.
- Fluid Animation: Fluid-animation methods use neural networks with fluid equations or reusable space-time regions to synthesize realistic simulations.These studies primarily target animation and simulation acceleration rather than the paper’s turbulent-flow forecasting problem.
- Video Prediction: Video-prediction models forecast future frames from observed frames but generally model natural videos with unknown physical processes.The paper’s turbulent-flow task predicts fluid evolution rather than object or camera motion.
3 BACKGROUND IN TURBULENCE MODELING
Turbulent flows are difficult to predict because they are chaotic, intermittent, and span many correlated length and time scales. CFD methods therefore resolve larger scales while modeling smaller-scale processes, while deep learning offers speed but can lack physical knowledge.
- Turbulent fluctuations span wide, correlated length and time scales, producing chaotic and intermittent motions that are difficult to predict.
- Two-dimensional Rayleigh-Bénard convection models turbulent convection by heating a horizontal fluid layer from below.
- The RBC simulation uses a Boussinesq approximation that produces a divergence-free flow, requiring ∇· w to equal zero everywhere.
- Fully resolving turbulent flows with DNS requires fine space-time discretization and is computationally prohibitive.
- RANS and LES resolve large scales while modeling small scales, but unresolved-scale interactions remain extremely challenging to represent.
- Deep learning can generate realistic flow fields without solving coupled nonlinear PDEs, but its lack of underlying physics makes models difficult to interpret.
4 METHODOLOGY
The methodology combines multiscale turbulence modeling with trainable deep-learning modules. TF-Net separates turbulent-flow scales, processes them with learned filters and encoders, and decodes their interactions into forecasts while optionally enforcing zero divergence.
- Computational Fluid Dynamics: RANS uses time averaging for simpler steady equations, whereas LES uses spatial filtering to improve accuracy at higher computational cost.
- Computational Fluid Dynamics: Hybrid RANS-LES coupling decomposes the flow into mean, resolved-fluctuation, and unresolved subgrid components using sequential temporal averaging and spatial filtering.
- Turbulent Flow Net: TF-Net replaces predefined RANS-LES spectral filters with trainable convolutional layers for multilevel spectral decomposition.
- Turbulent Flow Net: TF-Net uses trainable 5×5 spatial and 1×1 temporal convolutional filters to separate three velocity components at different scales.
- Turbulent Flow Net: Three identical encoders process the scale components separately, while a shared decoder learns their interactions and generates the next 2D velocity field.
- Turbulent Flow Net: Multi-step forecasting rolls one-step predictions forward autoregressively, with Con TF-Net adding ||∇· w||2 to the loss to enforce zero divergence.
5 EXPERIMENTS
Experiments evaluate TF-Net on simulated two-dimensional turbulent velocity fields against data-driven and physics-informed baselines using pixel and physics-based metrics. TF-Net consistently improves long-horizon accuracy, physical consistency, spectral fidelity, visualization quality, and generalization, while divergence regularization introduces a trade-off with RMSE.
- Dataset and baselines: The experiments use 1,500 two-dimensional turbulent-flow images with 1792 × 256 resolution and two velocity channels, generating 9,870 sequence samples.The samples comprise 6,000 training, 1,700 validation, and 2,170 test sequences from 256 × 256 sub-regions downsampled to 64 × 64.
- Dataset and baselines: TF-Net is compared with ResNet, ConvLSTM, U-Net, GAN, SST, and DHPM baselines for turbulent-flow prediction.ResNet, ConvLSTM, U-Net, and GAN are data-driven models, while SST and DHPM incorporate physical knowledge.
- Accuracy and efficiency: RMSE grows with forecasting horizon to 60 time steps, but TF-Net consistently outperforms all baselines.A divergence-free regularizer can further improve performance, whereas DHPM overfits training data and performs poorly outside the training domain.
- Physical metrics: TF-Net has lower prediction divergence than other models, although excessive divergence-free regularization smooths small eddies and increases RMSE.The results expose a trade-off between pixelwise accuracy and preservation of small-scale flow structure.
- Physical metrics: TF-Net predictions are closer to the target energy spectrum than U-Net and ResNet, despite similar turbulence kinetic-energy fields.The additional divergence-free constraint does not affect the prediction energy spectrum.
- Prediction visualization: TF-Net produces predictions closest to the target in motion shape and frequency, while baselines generate smoother fields that miss small-scale motion.All models still have room for improvement in long-term prediction.
- Multi-scale behavior: TF-Net’s multi-scale outputs indicate that its small U-nets learn flow behavior at different scales.The learned filters and zeroed-encoder ablations support the model’s scale-separated design.
- Generalization capability: On an additional dataset with Rayleigh number = 10^5, TF-Net outperforms U-Net and ResNet across RMSE, divergence, energy spectrum, and turbulence kinetic energy metrics.This result demonstrates generalization to turbulent flows with a different Rayleigh number.
6 DISCUSSION AND FUTURE WORK
The paper combines turbulence-simulation paradigms with deep learning in TF-Net and evaluates predictions using both accuracy and physical-consistency measures. It reports improved prediction errors and physical quantities, while identifying extensions to higher-resolution flows and additional variables and constraints.
- TF-Net unifies representation learning with turbulence simulation techniques through trainable scale-separation operators.
- The authors report significant improvement over baselines in prediction error, divergence, turbulence kinetic energy, and energy spectrum.
- Future work targets very high-resolution 3D turbulent flows, additional physical variables, and constraints such as momentum conservation.
Appendix
The appendix presents visual comparisons of predicted velocity fields and model performance, alongside the hyperparameter ranges used for tuning.
- Figure 12 compares ground-truth and predicted v velocities by model at forecasting times indexed relative to the last input frame.
- Figure 13 compares TF-Net, U-net, and ResNet on an additional dataset with Ra = 10000 across error, divergence, energy spectrum, and turbulence kinetic energy metrics.
- Table 2 lists tuning ranges for learning rate, batch size, accumulated backpropagation errors, input frames, moving-average window, and spatial filter size.