Towards Physics-informed Deep Learning for Turbulent Flow Prediction

Rui Wang, Karthik Kashinath, Mustafa Mustafa, Adrian Albert, Rose Yu

Abstract

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. 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, called turbulent-Flow Net (TF-Net), is grounded in a principled physics model yet offers the flexibility of learned representations. We observe significant reductions in error for predictions 60 frames ahead compared to state-of-the-art baselines. 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.

Introduction

Predicting turbulent flow is a fundamental challenge in classical physics due to its chaotic, multi-scale nature. Traditional Computational Fluid Dynamics (CFD) methods are physically accurate but computationally expensive. Conversely, pure deep learning (DL) models offer speed but often produce "unphysical" results that violate conservation laws. This research introduces TF-Net (Turbulent-Flow Net), a hybrid architecture designed to bridge this gap by embedding classical fluid dynamics principles—specifically scale separation—directly into a neural network.

Methodology

The core innovation of TF-Net is the integration of a popular CFD technique: RANS-LES coupling. In traditional physics, Reynolds-Averaged Navier-Stokes (RANS) handles large-scale mean flow, while Large Eddy Simulation (LES) focuses on resolved fluctuations.

Decomposition of the energy spectrum into mean flow, resolved fluctuations, and unresolved fluctuations.

TF-Net translates these concepts into a deep learning framework using:

  1. Trainable Spectral Filters: Instead of fixed mathematical filters, TF-Net uses convolutional layers to learn how to decompose the flow into three components: mean flow, resolved fluctuations, and unresolved fluctuations.
  2. Multi-Scale Encoders: Three identical encoders process these scales separately, allowing the network to learn scale-specific dynamics.
  3. Shared Decoder: A U-Net-style decoder then recombines these features to predict the future state of the fluid.
  4. Physical Constraints: A "divergence-free" regularizer is added to the loss function to enforce the conservation of mass.
The TF-Net architecture, illustrating the flow from spatial/temporal filters through multi-scale encoders to a shared U-Net decoder.

The model was trained on Rayleigh-Bénard Convection (RBC) data, generated via the Lattice Boltzmann Method (LBM). The task involved predicting 60 frames of fluid motion given a 10-frame history of high-resolution velocity fields.

Results and Insights

TF-Net demonstrates superior performance over data-driven baselines like ResNet, ConvLSTM, and GANs:

  • Accuracy: Achieved an 11.1% reduction in RMSE compared to the best baseline.
  • Physical Fidelity: Showed a 30.1% improvement in the energy spectrum and a 64.2% reduction in flow divergence, indicating much better adherence to the conservation of mass.
  • Scale-Aware Learning: By forcing the network to analyze different scales separately, the model learns underlying dynamics more efficiently than a single monolithic network.
  • Constraint Trade-offs: The study identified a trade-off where heavy physical constraints can improve consistency but may slightly smooth out fine-scale chaotic features, increasing pixel-wise error.

Conclusion

TF-Net represents a significant step toward physics-informed deep learning for fluid dynamics. By making physical filters trainable, the model adapts to specific datasets while maintaining physical rigor. Future work will extend this approach to 3D flows and incorporate additional physical variables like pressure and temperature to further enhance long-term prediction stability in chaotic systems.

Summary updated

Abstract

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.
BibTeX
@inproceedings{wang2020towards,
  author = {Rui Wang and Karthik Kashinath and Mustafa Mustafa and Adrian Albert and Rose Yu},
  title = {{Towards Physics-informed Deep Learning for Turbulent Flow Prediction}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2020},
  doi = {10.1145/3394486.3403198},
}
Figure