A dynamic subgrid scale model for Large Eddy Simulations based on the Mori-Zwanzig formalism

Eric J. Parish, Karthik DuraisamyJournal of Computational Physics 2017 · OpenReview

TL;DR

This paper introduces a novel, parameter-free subgrid-scale model for Large Eddy Simulations (LES) called the "dynamic-MZ-τ" model. It leverages the Mori-Zwanzig formalism to mathematically derive a closure that captures non-local memory effects in turbulent flows, outperforming existing models in certain scenarios and offering a more rigorous, physics-driven approach.

What Problem Does This Paper Address?

Computational physics often grapples with multiscale problems, where phenomena occur across vastly different spatial and temporal scales. Direct numerical simulations (DNS) that resolve all scales are computationally prohibitive for many real-world applications, especially turbulence. Large Eddy Simulations (LES) offer a compromise by filtering out small scales and modeling their effects. However, the process of coarse-graining in LES can introduce "memory effects" – non-local influences of unresolved scales on resolved ones – which are challenging to quantify and model. Traditional subgrid-scale models often rely on phenomenological assumptions, lacking a strong theoretical foundation. This paper aims to address this by developing a mathematically derived, parameter-free LES closure model based on the Mori-Zwanzig formalism, which explicitly accounts for these memory effects.

Methods

The core of this work lies in the application and extension of the Mori-Zwanzig (M-Z) formalism. This framework allows for the transformation of a high-dimensional, Markovian (memoryless) dynamical system into a lower-dimensional, non-Markovian (memory-dependent) system. The key steps involve:

  • Generalized Langevin Equation (GLE): The M-Z formalism leads to a GLE, which expresses the dynamics of resolved variables in terms of a convolution integral representing the influence of unresolved scales.
  • Orthogonal Dynamics: To evaluate this convolution integral, the paper introduces the concept of "orthogonal dynamics," which describes the behavior of the unresolved scales.
  • Approximation of Memory: Directly solving the orthogonal dynamics is intractable. The paper proposes approximating the memory integral by assuming it has a finite temporal support (a "memory length," denoted by $\tau_p$).
  • Germano Identity: To dynamically determine this memory length $\tau_p$, the authors employ the Germano identity, a concept from subgrid-scale modeling, to establish a relationship between different levels of coarse-graining.
  • Dynamic Procedure: This leads to a "dynamic-t" model where the memory length $\tau_p$ is computed on-the-fly, making the model parameter-free. The derivation involves analytical evaluation of certain terms and numerical approximations for others.

Datasets

The paper uses data from:

  • Direct Numerical Simulations (DNS): The authors compare their model's performance against DNS data for:
  • Burgers equation
    • Homogeneous isotropic turbulence at various Reynolds numbers (low, moderate, and high)
    • Turbulent channel flow
  • No new benchmarks were introduced. The focus was on validating the proposed model against existing, well-established DNS datasets.

Key Insights and Ideas

  • Mathematically Imposed Structure: The "dynamic-MZ-τ" model's structure is derived from the mathematics of coarse-graining and the Mori-Zwanzig formalism, rather than relying on ad-hoc phenomenological assumptions common in traditional LES models. This offers a more rigorous foundation.
  • Finite Memory Support: The assumption that memory effects have a finite temporal support is crucial and is supported by numerical evidence, particularly for decaying turbulence. This allows for a more tractable model.
  • Dynamic Memory Length: By using the Germano identity, the model dynamically determines the memory length ($\tau_p$), making it parameter-free and adaptable to different flow conditions and resolutions. This is a significant advantage over models that require manual tuning of such parameters.
  • Predicting Backscatter: For homogeneous turbulence, the dynamic-MZ-τ model shows an ability to predict "backscatter" – the transfer of energy from small scales to larger scales – which is a physically important phenomenon often missed by simpler models.
  • Channel Flow Behavior: The model's ability to predict a constant, rather than growing, memory time-scale for statistically steady flows like channel flow is a key differentiator from models that assume a growing memory length.

Key Results

  • Burgers Equation: The dynamic-MZ-τ model accurately predicts the decay of kinetic energy, subgrid content, and velocity field trajectories, comparing favorably to DNS and outperforming LES with no model.
  • Homogeneous Turbulence:
    • At low Reynolds numbers, the model shows excellent agreement with DNS for energy decay and spectra, and captures subgrid energy transfer well. It also demonstrates backscatter prediction.
  • At moderate and high Reynolds numbers, the model's performance slightly degrades, particularly in predicting energy dissipation. However, it still generally outperforms LES with no model and, in some cases, the dynamic Smagorinsky model.
  • Turbulent Channel Flow: The model provides reasonable predictions for mean velocity profiles and Reynolds stresses, and importantly, predicts a time-scale that is consistent with the steady-state nature of the flow.
  • Parameter-Free: The model is demonstrated to be parameter-free, a significant achievement for LES closure modeling.

Limitations and Future Work

  • Performance Degradation at High Reynolds Numbers/Low Resolution: The paper acknowledges that the model's performance degrades as the LES becomes increasingly coarse or for higher Reynolds numbers. This is attributed to the breakdown of the assumption that the memory integral is well-correlated to its value at time zero, as the orthogonal dynamics become more complex.
  • Computational Cost: While the model is parameter-free, the evaluation of certain terms, particularly the Fréchet derivative, can increase computational cost compared to simpler models.
  • Future Work:
    • Developing higher-order approximations for the memory integral.
  • Exploring dynamic procedures for hierarchical finite memory models.
  • Extending the model to wall-bounded turbulence with more sophisticated treatments of boundary conditions.
  • Investigating the model within variational multiscale (VMS) frameworks, which share similarities with the M-Z approach.

Summary updated

Abstract

The development of reduced models for complex multiscale problems remains one of the principal challenges in computational physics. The optimal prediction framework of Chorin et al. [1], which is a reformulation of the Mori–Zwanzig (M–Z) formalism of non-equilibrium statistical mechanics, provides a framework for the development of mathematically-derived reduced models of dynamical systems. Several promising models have emerged from the optimal prediction community and have found application in molecular dynamics and turbulent flows. In this work, a new M–Z-based closure model that addresses some of the deficiencies of existing methods is developed. The model is constructed by exploiting similarities between two levels of coarse-graining via the Germano identity of fluid mechanics and by assuming that memory effects have a finite temporal support. The appeal of the proposed model, which will be referred to as the ‘dynamic-MZ-τ’ model, is that it is parameter-free and has a structural form imposed by the mathematics of the coarse-graining process (rather than the phenomenological assumptions made by the modeler, such as in classical subgrid scale models). To promote the applicability of M–Z models in general, two procedures are presented to compute the resulting model form, helping to bypass the tedious error-prone algebra that has proven to be a hindrance to the construction of M–Z-based models for complex dynamical systems. While the new formulation is applicable to the solution of general partial differential equations, demonstrations are presented in the context of Large Eddy Simulation closures for the Burgers equation, decaying homogeneous turbulence, and turbulent channel flow. The performance of the model and validity of the underlying assumptions are investigated in detail.
BibTeX
@article{parish2017dynamic,
  author = {Eric J. Parish and Karthik Duraisamy},
  title = {{A dynamic subgrid scale model for Large Eddy Simulations based on the Mori-Zwanzig formalism}},
  journal = {Journal of Computational Physics},
  year = {2017},
  url = {https://openreview.net/forum?id=7lkH8AaNI2},
}
Figure