Towards Physics-informed Deep Learning for Turbulent Flow Prediction

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

TL;DR

Researchers have developed TF-Net (Turbulent-Flow Net), a hybrid deep learning architecture that embeds classical fluid dynamics principles—specifically scale separation—directly into a neural network. By decomposing chaotic flows into different scales using trainable filters, TF-Net achieves state-of-the-art prediction accuracy while ensuring the results obey physical laws like the conservation of mass.

What Problem Does This Paper Address?

Predicting turbulent flow is often called the "last great problem of classical physics." It is chaotic, multi-scale (small eddies interacting with large waves), and computationally grueling.

Currently, we have two main tools, both with flaws:

  1. Computational Fluid Dynamics (CFD): These are purely physics-based simulations. They are highly accurate but incredibly slow and expensive, requiring massive supercomputers to solve complex partial differential equations (PDEs).
  2. Pure Deep Learning (DL): These models are fast and great at pattern recognition, but they are "black boxes." They often produce "unphysical" results—like fluids gaining mass out of nowhere—because they don't "know" the laws of physics.

The authors aim to close this gap by creating a model that is as fast as a neural network but as physically consistent as a CFD simulation.

Methods

The core innovation is TF-Net, which mimics a popular CFD technique called RANS-LES coupling.

In traditional CFD:

  • RANS (Reynolds-Averaged Navier-Stokes) handles the "big picture" by averaging flow over time.
  • LES (Large Eddy Simulation) uses spatial filters to focus on larger, more energetic structures.

TF-Net translates these concepts into a deep learning architecture:

  • Trainable Spectral Filters: Instead of using fixed mathematical filters, TF-Net uses convolutional layers to learn how to best separate the flow into three components: the mean flow (large scale), resolved fluctuations (medium scale), and unresolved fluctuations (small scale).
  • Multi-Scale Encoders: The model uses three identical encoders to process these three scales separately.
  • Shared Decoder: A U-Net-style decoder then recombines these features to predict the next state of the fluid.
  • Physical Constraints: The authors added a "divergence-free" regularizer to the loss function. In simple terms, this forces the model to ensure that the amount of fluid flowing into a point equals the amount flowing out (conservation of mass).

Datasets

The model was trained and evaluated on Rayleigh-Bénard Convection (RBC) data. This is a classic physics setup where a fluid is heated from below and cooled from above, creating turbulent "cells." * Simulation: Data was generated using the Lattice Boltzmann Method (LBM). * Scale: The dataset consisted of high-resolution (1792 x 256) velocity fields. * Task: Given 10 frames of history, predict the next 60 frames of fluid motion.

Key Insights and Ideas

  • Scale Matters: The "aha moment" here is that a single neural network shouldn't try to learn everything at once. By forcing the network to look at different scales separately (just like physicists do), the model learns the underlying dynamics much more efficiently.
  • Trainable vs. Fixed: In traditional physics, filters are pre-defined. By making these filters trainable parameters of the neural network, the model can adapt its "physics" to the specific dataset it's looking at.
  • The Accuracy-Consistency Trade-off: The researchers found a subtle trade-off. Adding strict physical constraints (like the divergence-free regularizer) makes the model more physically "correct," but if the constraint is too heavy, it can slightly smooth out the tiny, chaotic eddies, leading to a higher pixel-wise error (RMSE).

Key Results

TF-Net significantly outperformed pure data-driven baselines (like ResNet, ConvLSTM, and GANs) across several metrics:

  • Accuracy: Achieved an 11.1% reduction in RMSE compared to the best baseline.
  • Physical Fidelity: Showed a 30.1% improvement in the energy spectrum (how well it captures eddies of different sizes) and a 64.2% reduction in flow divergence (meaning it respects the conservation of mass far better than standard models).
  • Efficiency: Once trained, TF-Net is significantly faster than traditional numerical solvers, though the authors note that highly optimized LBM solvers are already quite efficient.

Limitations and Future Work

While a major step forward, the authors acknowledge several areas for growth:

  • 2D vs. 3D: This study focused on 2D flows. Real-world turbulence is 3D, which adds a massive layer of complexity and data requirements.
  • Variable Complexity: The current model predicts velocity but doesn't yet incorporate other critical physical variables like pressure and temperature into the constraints.
  • Long-term Chaos: Like all weather-related predictions, the chaotic nature of turbulence means that even with physics-informed models, accuracy eventually degrades over very long time horizons.

Future work will look into extending these hybrid techniques to 3D simulations and incorporating more complex physical laws, such as the conservation of momentum.

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, is grounded in a principled physics model, yet offers the flexibility of learned representations. We compare our model 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},
}

Loading related content...

Figure