Automated Dissipation Control for Turbulence Simulation with Shell Models

Ann-Kathrin Dombrowski, Klaus-Robert Müller, Wolf-Christian MüllerarXiv 2022 · arXiv

Introduction

Simulating turbulent fluid flow remains one of the most significant challenges in classical physics. While Direct Numerical Simulation (DNS) provides high-fidelity results by calculating every eddy and swirl, its computational cost is prohibitive for practical applications like meteorology or aerospace engineering. Consequently, researchers rely on Large-Eddy Simulations (LES), which resolve large-scale structures and employ subgrid-scale (SGS) models to approximate small-scale dynamics.

Traditional machine learning (ML) approaches to SGS modeling often involve training neural networks to mimic high-resolution DNS data. However, these models frequently suffer from long-term instability; minor errors accumulate, leading to unphysical "explosions" or drift. This paper proposes a "learning-by-evolution" framework that embeds ML directly into a simplified physical system, tuning the model to satisfy fundamental statistical laws rather than merely replicating snapshots of data.

Methodology: The GOY Shell Model and Physics-Constrained Learning

The researchers utilize the Gledzer-Ohkitani-Yamada (GOY) shell model, a one-dimensional simplification of turbulence. This model captures the essential "energy cascade"—the process where energy transfers from large-scale eddies to smaller ones until it is eventually dissipated by viscosity.

Transition from 2D real velocity fields to the 1D complex velocity representation used in shell models.

In this framework, the standard dissipation (viscosity) term is replaced with a learnable parameter, $\theta$. The objective is for the ML algorithm to autonomously discover the correct dissipation required to stabilize the system. Unlike standard supervised learning, the loss function here is statistical: it measures the deviation of the simulation's energy spectrum from Kolmogorov’s $-5/3$ law.

Schematic of the energy cascade showing the energy-containing, inertial, and dissipation ranges.

As shown in the figure above, the energy spectrum follows a specific power-law slope in the "inertial subrange." By using this spectrum as a "statistical sensor," the model receives feedback on whether its dynamics are physically consistent. If the slope deviates from $-5/3$, the model adjusts $\theta$ to correct the dissipation.

Technical Implementation

To optimize the parameter $\theta$, the authors employ the adjoint method. This technique allows for the calculation of gradients by solving an auxiliary equation backwards in time, avoiding the massive memory overhead associated with storing every intermediate step of a chaotic simulation. The system was implemented in the Julia programming language to leverage its efficient handling of complex numbers and advanced differential equation solvers.

Results and Discussion

The "learning-by-evolution" approach successfully demonstrated that an ML model can "relearn" physical constants from scratch. Starting with zero initial knowledge of dissipation, the parameter $\theta$ converged and oscillated around the true physical value of $10^{-8}$.

Key findings include:

  • Statistical Fidelity: The simulation accurately recovered the characteristic $-5/3$ slope in the energy spectrum and maintained a constant energy flux across the inertial range.
  • Stability through Statistics: By targeting long-term statistical properties rather than individual data points, the model achieved inherent robustness, avoiding the common "explosion" issues seen in data-mimicry models.
  • Timescale Sensitivity: The study highlighted that the ML update frequency must align with the "large-eddy turnover time" of the physics. Updating too slowly prevents the model from reacting to chaotic shifts, while updating too quickly provides insufficient statistical data for meaningful learning.

Conclusion and Future Work

This work serves as a proof of concept for automated dissipation control in dynamical systems. While successful in the 1D GOY shell model, the next frontier involves scaling this approach to the 3D Navier-Stokes equations. Future challenges include managing the increased computational complexity and implementing more sophisticated "hard" physical constraints within the neural architecture to prevent unphysical parameter regimes, such as negative viscosity.

Summary updated

BibTeX
@misc{dombrowski2022automated,
  author = {Ann-Kathrin Dombrowski and Klaus-Robert Müller and Wolf-Christian Müller},
  title = {{Automated Dissipation Control for Turbulence Simulation with Shell Models}},
  year = {2022},
  eprint = {2201.02485},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2201.02485},
}
Figure