Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learning

Yifei Guan, Ashesh Chattopadhyay, Adam Subel, Pedram HassanzadehJournal of Computational Physics 2022 · DOI

TL;DR

This paper introduces a novel deep convolutional neural network (CNN) approach for modeling subgrid-scale (SGS) turbulence in large-eddy simulations (LES). By leveraging transfer learning, their method achieves stable and accurate LES simulations across a wide range of Reynolds numbers, overcoming key limitations of previous data-driven turbulence models.

What Problem Does This Paper Address?

Accurately simulating turbulent flows is crucial for many scientific and engineering applications. Direct numerical simulations (DNS) are too computationally expensive for most real-world problems. Large-eddy simulation (LES) offers a compromise by resolving large-scale turbulent structures and parameterizing the effects of smaller, subgrid scales (SGS) using an SGS model. However, developing accurate and stable SGS models remains a significant challenge. Traditional physics-based models often struggle to capture complex turbulent phenomena like "backscattering" (energy transfer from small to large scales), which can lead to instabilities in LES. Data-driven approaches using machine learning (ML) have shown promise in capturing these effects, but often suffer from instabilities and poor generalization to flows with different parameters (e.g., higher Reynolds numbers, Re). This paper aims to address these issues by developing a robust, generalizable, and stable data-driven SGS model.

Methods

The authors propose a non-local, data-driven SGS model based on a deep fully convolutional neural network (CNN). Unlike local models that consider only nearby grid points, this CNN takes the entire domain's resolved flow variables (vorticity and streamfunction) as input to capture spatial correlations. The model is trained to predict the SGS forcing term, which represents the effect of subgrid scales on the resolved flow.

A key innovation is the application of transfer learning. This technique involves training a base CNN on data from a low-Reynolds number flow and then "fine-tuning" it with a small amount of data from a higher-Reynolds number flow. This allows the model to generalize to new flow conditions without requiring extensive retraining. For handling different LES grid resolutions, an encoder-decoder architecture is incorporated into the CNN.

Datasets

The study utilizes data from high-fidelity direct numerical simulations (DNS) of 2D decaying homogeneous isotropic turbulence. These DNS runs are performed at various Reynolds numbers (Re = 8000, 32000, 64000, and 128000) and resolutions. The DNS data is filtered to create filtered DNS (FDNS) data, which serves as the ground truth for training and testing the SGS models. The training dataset consists of snapshots of filtered vorticity and streamfunction, and the target is the corresponding SGS forcing term.

Key Insights and Ideas

  • Backscattering is Crucial for Stability: The paper strongly suggests that accurate prediction of backscattering is essential for achieving stable LES simulations. Models that fail to capture this phenomenon, or actively remove it, tend to be overly dissipative and lead to inaccurate results or instabilities.
  • Transfer Learning is a Game-Changer for Generalization: The effectiveness of transfer learning in enabling data-driven SGS models to generalize to significantly higher Reynolds numbers (up to 16x) is a major breakthrough. This addresses a critical bottleneck for practical applications of ML in turbulence modeling.
  • CNNs Outperform ANNs in Capturing Spatial Structures: The study reinforces the idea that CNNs are better suited than traditional ANNs for turbulence modeling due to their ability to inherently capture spatial patterns and correlations, which are vital in turbulent flows.
  • Sufficient Training Data is Key to Stability: While a priori accuracy metrics (like correlation coefficients) might look good with limited training data, a posteriori LES stability is only achieved when the training dataset is sufficiently large, particularly for accurately learning backscattering.

Key Results

  • High Accuracy in A Priori Tests: The CNN-based SGS model achieved a correlation coefficient (c) of around 0.93 in predicting the SGS forcing term, significantly outperforming physics-based models like DSMAG (c ≈ 0.55).
  • Stable and Accurate A Posteriori LES: With sufficient training data (ntr ≥ 30000), the LES-CNN simulations were stable and accurately reproduced the kinetic energy decay rate and vorticity statistics, closely matching FDNS results.
  • Effective Generalization via Transfer Learning: The transfer-learned CNN (LES-TL-CNN) successfully extrapolated to 4x and 8x higher Reynolds numbers, achieving accuracy comparable to models trained directly on high-Re data. This was achieved with only 1% of the original training data for fine-tuning.
  • Handling Higher Resolutions: The integration of an encoder-decoder architecture allowed the transfer-learned model to also adapt to higher LES grid resolutions.
  • Outperforming Other Models: In a posteriori tests, LES-CNN demonstrated superior short-term prediction accuracy and long-term statistical reproduction compared to DSMAG and a post-processed LES-ANN.

Limitations and Future Work

  • Data Requirements for Backscattering: The paper notes that accurately learning backscattering might require more training data than forward transfer, and the exact reasons for this are still an area for future study.
  • Generalization to More Complex Flows: While successful for 2D decaying turbulence, the authors suggest exploring transfer learning and CNNs for more complex flows, such as 3D, wall-bounded turbulence, or stratified flows.
  • Reducing Training Data: Future work could investigate methods like data augmentation (exploiting symmetries) or physics-informed ML to potentially reduce the amount of training data needed.
  • Interpreting CNNs: Understanding the physical insights gained from the trained CNNs is an ongoing challenge and a promising avenue for future research.
  • Online Training: The authors mention that an "online training" strategy, where the model learns and evolves with the simulation, could potentially resolve some stability issues and improve scaling.

Summary updated

BibTeX
@article{guan2022stable,
  author = {Yifei Guan and Ashesh Chattopadhyay and Adam Subel and Pedram Hassanzadeh},
  title = {{Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learning}},
  journal = {Journal of Computational Physics},
  year = {2022},
  doi = {10.1016/j.jcp.2022.111090},
}
Figure