Data-Driven Mori-Zwanzig: Approaching a Reduced Order Model for Hypersonic Boundary Layer Transition

Mark Woodward, Yifeng Tian, Arvind Mohan, Yen Ting Lin, Christoph Hader, Hermann F. Fasel, Michael Chertkov, Daniel LivescuAIAA 2023 · arXiv

TL;DR

Researchers have applied the Mori-Zwanzig (MZ) formalism to hypersonic boundary layer transitions, creating a reduced-order model that outperforms standard Dynamic Mode Decomposition (DMD). By accounting for "memory effects"—how past states influence future ones—this approach improves prediction accuracy by up to 15% and identifies physical structures like "hot streaks" that Markovian models often miss.

What Problem Does This Paper Address?

Predicting how air flows over a vehicle at Mach 6 is a monumental challenge. As the flow transitions from smooth (laminar) to chaotic (turbulent), it creates "hot streaks" on the vehicle's surface, leading to massive increases in drag and heat.

Engineers use Direct Numerical Simulations (DNS) to model this, but DNS is computationally "expensive"—it takes massive supercomputing power to resolve every tiny swirl. To save time, researchers use Reduced Order Models (ROMs) like Dynamic Mode Decomposition (DMD). However, DMD is "Markovian": it assumes the next state of the flow depends only on the current state. In complex, nonlinear hypersonic transitions, the flow has a "memory" where unresolved small-scale dynamics continue to influence large-scale behavior over time. This paper seeks to bridge that gap by capturing those non-Markovian memory effects.

Methods

The authors utilize the Mori-Zwanzig (MZ) formalism, a concept from statistical mechanics. MZ mathematically splits a high-dimensional system into three parts:

  1. Markovian Term: The immediate, linear evolution (equivalent to DMD).
  2. Memory Term: A "kernel" that accounts for the history of the resolved variables.
  3. Noise Term: The influence of unresolved, orthogonal dynamics (often treated as a small residual).

The team developed a data-driven algorithm that uses Singular Value Decomposition (SVD) to compress high-dimensional DNS data into a set of "resolved variables." They then learn the Markovian and memory operators directly from this data. This allows them to build a Generalized Langevin Equation (GLE) that can predict future flow states more accurately than traditional linear methods.

Datasets

The study evaluates the algorithm on two primary datasets:

  • 2D Flow over a Cylinder (Re=100): A standard benchmark used for initial validation to ensure the MZ Markovian terms match known DMD results.
  • Hypersonic Flared Cone (Mach 6): High-fidelity DNS data of a boundary layer transition. This is a spatially inhomogeneous flow, making it a much more complex test case than previous MZ applications.

Key Insights and Ideas

  • DMD is a Subset of MZ: The authors prove that the Markovian operator in their MZ framework is identical to the best-fit operator found in DMD. This means MZ is essentially "DMD plus memory."
  • Physicality of Memory: The "memory kernels" aren't just mathematical corrections; they correspond to real physical structures. Specifically, the memory terms identified the nonlinear generation of "hot streaks" on the cone's surface—features that are critical for engineering but difficult to capture with linear models.
  • Inverse Cascade: While the Markovian terms capture the cascade from low-frequency instabilities to higher harmonics, the memory terms demonstrate an "inverse cascade" effect, helping to resolve the transition region's complex physics.

Key Results

  • Improved Accuracy: By including memory terms, the MZ model achieved a 5% to 15% improvement in prediction accuracy over DMD at the same level of truncation.
  • Better Feature Resolution: In the transition region of the flared cone, DMD predictions showed larger errors, particularly around the "hot streaks." The MZ approach significantly reduced these errors by accounting for the hysteresis (history-dependent) effects of the flow.
  • Computational Efficiency: The MZ algorithm provides these higher-order corrections at a computational cost similar to DMD, making it a viable candidate for real-time engineering applications.

Limitations and Future Work

The authors acknowledge a few areas for growth:

  • The Noise Term ($F$): In this study, the "noise" or orthogonal dynamics term was assumed to be negligible during prediction. Future work could use nonlinear projection or regression to better model this residual.
  • 3D Structures: The current analysis focused on surface pressure; extending this to extract 3D coherent structures from volumetric data is a logical next step.
  • Observable Selection: The choice of "resolved variables" (observables) is crucial. The authors suggest investigating different ways to select these variables to further optimize the model's performance.

Summary updated

Abstract

In this work, we apply, for the first time to spatially inhomogeneous flows, a recently developed data-driven learning algorithm of Mori-Zwanzig (MZ) operators, which is based on a generalized Koopman's description of dynamical systems. The MZ formalism provides a mathematically exact procedure for constructing non-Markovian reduced-order models of resolved variables from high-dimensional dynamical systems, where the effects due to the unresolved dynamics are captured in the memory kernel and orthogonal dynamics. The algorithm developed in this work applies Mori's linear projection operator and an SVD based compression to the selection of the resolved variables (equivalently, a low rank approximation of the two time covariance matrices). We show that this MZ decomposition not only identifies the same spatio-temporal structures found by DMD, but it can also be used to extract spatio-temporal structures of the hysteresis effects present in the memory kernels. We perform an analysis of these structures in the context of a laminar-turbulent boundary-layer transition flow over a flared cone at Mach 6, and show the dynamical relevance of the memory kernels. Additionally, by including these memory terms learned in our data-driven MZ approach, we show improvement in prediction accuracy over DMD at the same level of truncation and at a similar computational cost. Furthermore, an analysis of the spatio-temporal structures of the MZ operators shows identifiable structures associated with the nonlinear generation of the so-called "hot" streaks on the surface of the flared code, which have previously been observed in experiments and direct numerical simulations.
BibTeX
@inproceedings{woodward2023data,
  author = {Mark Woodward and Yifeng Tian and Arvind Mohan and Yen Ting Lin and Christoph Hader and Hermann F. Fasel and Michael Chertkov and Daniel Livescu},
  title = {{Data-Driven Mori-Zwanzig: Approaching a Reduced Order Model for Hypersonic Boundary Layer Transition}},
  booktitle = {AIAA},
  year = {2023},
}
Figure