Data-Driven Mori-Zwanzig: Reduced Order Modeling of Sparse Sensors Measurements for Boundary Layer Transition

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

TL;DR

Researchers have developed a data-driven framework using the Mori-Zwanzig (MZ) formalism to predict boundary-layer transition—the chaotic shift from smooth to turbulent flow—using only a handful of surface sensors. By mathematically accounting for "memory effects" (how unobserved parts of the flow influence the sensors over time), their regression-based MZ model outperforms traditional methods like EDMD and rivals deep learning models like LSTMs while maintaining a more interpretable physical structure.

What Problem Does This Paper Address?

In aerospace engineering, understanding when and where a flow turns turbulent (boundary-layer transition) is a "holy grail" problem. Turbulence drastically increases skin-friction drag and heat transfer, which dictates everything from fuel efficiency to the thickness of a spacecraft's thermal protection system.

The catch? In real flight tests or wind tunnels, we can’t see the whole flow field. We only have "sparse" data—pressure sensors at a few fixed points on the wing or fuselage. Standard models often struggle because they assume the system is Markovian (the future depends only on the current state). In reality, because we aren't measuring the entire fluid volume, the unobserved variables interact with our sensors in complex ways. These interactions manifest as "memory," where the history of the sensor readings contains vital clues about the hidden dynamics of the surrounding fluid.

Methods

The authors leverage the Mori-Zwanzig (MZ) formalism, a concept from statistical mechanics, to build Reduced Order Models (ROMs). The MZ procedure produces the Generalized Langevin Equation (GLE), which decomposes the evolution of observed variables into three parts:

  1. Markovian term: The immediate influence of the current state.
  2. Memory kernel: A term that integrates the history of the sensors to account for interactions with unobserved variables.
  3. Orthogonal dynamics: The "noise" or residual effects of the full system.

The paper compares several data-driven flavors of this approach:

  • Mori-based MZ: Uses linear projections to find the memory kernels.
  • Regression-based MZ: Uses fully connected Neural Networks (3 hidden layers) to learn non-linear MZ operators.
  • Baselines: They compared these against Extended Dynamic Mode Decomposition (EDMD) with delay embeddings and Long Short-Term Memory (LSTM) networks.

Datasets

The models were trained and validated on high-fidelity Direct Numerical Simulation (DNS) data from two distinct scenarios:

  1. Incompressible Flow over a Flat Plate: Data sourced from the Johns Hopkins Turbulence Database (JHTDB) representing "bypass transition" (transition triggered by external turbulence).
  2. Hypersonic Compressible Flow (Mach 6): DNS of a flared cone at zero angle of attack, where transition occurs "naturally" via the breakdown of second-mode instability waves.

In both cases, an array of virtual "sensors" was placed near the solid boundary to mimic real-world experimental setups.

Key Insights and Ideas

  • Memory vs. Delay: While "delay embedding" (stacking past observations into a single vector) is a common trick in Koopman operator theory, the authors found that explicitly modeling the Memory Kernel via MZ is more effective for capturing long-term statistical properties.
  • The "Aha" Moment on Scales: If you want to match the exact, instantaneous pressure signal (short-term), delay embeddings help significantly. However, if you want to predict the statistical distribution of the flow (KL-Divergence) as it advects downstream, the MZ memory terms are the dominant factor for accuracy.
  • Structure Matters: Unlike a "black-box" LSTM, the regression-based MZ model is built on a formal mathematical projection. This makes it a "gray-box" model that is more interpretable and potentially easier to constrain with physical laws in the future.

Key Results

  • Top Performer: The Regression-based MZ model (using neural networks) achieved the lowest generalization error and best captured the power spectrum of the pressure signals.
  • Close Second: LSTMs performed nearly as well as the regression MZ model, but the authors note that LSTMs do not explicitly enforce the Generalized Fluctuation-Dissipation (GFD) theorem, a core tenet of the MZ formalism.
  • Linear Superiority: Even the linear Mori-based MZ outperformed standard EDMD. The results showed that adding memory terms reduces generalization error more effectively than simply increasing the number of delay embeddings in a Markovian model.
  • Downstream Robustness: On the JHTDB flat plate data, the MZ approach maintained superior predictive accuracy even as the flow moved from the laminar regime into the highly non-linear turbulent regime.

Limitations and Future Work

The primary limitation acknowledged is that this study ignored the orthogonal dynamics (the "noise" term in the GLE), assuming it to be a small residual. While this worked for the tested cases, highly chaotic regimes might require modeling this term as a stochastic process.

Future work aims to use these MZ-based reduced models for active flow control. Because these models are computationally "cheap" compared to DNS but more accurate than standard ROMs, they could potentially run in real-time on an aircraft's flight computer to delay the onset of turbulence, saving fuel and reducing heat loads.

Summary updated

Abstract

View Video Presentation: https://doi.org/10.2514/6.2023-4256.vid Understanding, predicting and controlling laminar-turbulent boundary-layer transition is crucial for the next generation aircraft design. However, in real flight experiments, or wind tunnel tests, often only sparse sensor measurements can be collected at fixed locations. Thus, in developing reduced models for predicting and controlling the flow at the sensor locations, the main challenge is in accounting for how the surrounding field of unobserved (or unresolved) variables interacts with the observed (or resolved) variables at the fixed sensor locations. This makes the Mori-Zwanzig (MZ) formalism a natural choice, as it results in the Generalized Langevin Equations which provides a mathematically sound framework for constructing non-Markovian reduced-order models that includes the effects the unresolved variables have on the resolved variables. These effects are captured in the so called memory kernel and orthogonal dynamics, which, when using Mori's linear projection, provides a higher order approximation to the finite rank Koopman approximation. In this work, we explore recently developed data-driven methods for extracting the MZ operators to two boundary layer flows obtained from DNS data; a low speed incompressible flow over a flat plate exhibiting bypass transition; and a high speed compressible flow over a flared cone at Mach 6 and zero angle of attack where transition was initiated using a broadband forcing approach ("natural" transition). In each case, an array of "sensors" are placed near the surface of the solid boundary, and the MZ operators are learned and the predictions are compared to the Extended Dynamic Mode Decomposition (EDMD), both using delay embedded coordinates. Further comparisons are made with Long Short-Term Memory (LSTM) and a regression based projection framework using neural networks for the MZ operators. First we compare the effects of including delay embedded coordinates with EDMD and Mori based MZ and provide evidence that using both memory and delay embedded coordinates minimizes generalization errors on the relevant time scales. Next, we provide numerical evidence that the data-driven regression based projection MZ model performs best with respect to the prediction accuracy (minimum generalization error) on the relevant time scales.
BibTeX
@misc{woodward2023data,
  author = {Mark Woodward and Yifeng Tian and Yen Ting Lin and Arvind Mohan and Christoph Hader and Hermann F. Fasel and Michael Chertkov and Daniel Livescu},
  title = {{Data-Driven Mori-Zwanzig: Reduced Order Modeling of Sparse Sensors Measurements for Boundary Layer Transition}},
  year = {2023},
  eprint = {2309.15864},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.15864},
}
Figure