Search

We present a robust and flexible optimization approach for dynamic mode decomposition analysis of data with complex dynamics and low signal-to-noise ratios. The approach borrows techniques and insights from the field of deep learning. Specifically, we employ automatic differentiation and stochastic gradient descent to compute eigenvalues, modes, and mode amplitudes simultaneously. The method allows embedding regularization or physical constraints into the operator definition. The optimization approach is applied to three examples of increasing complexity, the most challenging of which is an experimental dataset of transonic shock buffet on a swept at realistic flight conditions.

During the last decade, forcing and response modes produced by resolvent analysis have demonstrated great potential to guide sensor and actuator placement and design in flow control applications. However, resolvent modes are frequency-dependent, which, although responsible for their success in identifying scale interactions in turbulence, complicates their use for control purposes. In this work, we seek orthogonal bases of forcing and response modes that are the most responsive and receptive, respectively, across all frequencies. We show that these frequency-independent bases of \emph{representative} resolvent modes are given by the eigenvectors of the observability and controllability Gramians of the system considering full state inputs and outputs. We present several numerical examples where we leverage these bases by building orthogonal or interpolatory projectors onto the dominant forcing and response subspaces. Gramian-based forcing modes are used to identify dynamically relevant disturbances, to place point sensors to measure disturbances, and to design actuators for feedforward control in the subcritical linearized Ginzburg--Landau equation. Gramian-based response modes are used to identify coherent structures and for point sensor placement aiming at state reconstruction in the turbulent flow in a minimal channel at $\mathrm{Re}_{\tau}=185$. The approach does not require data snapshots and relies only on knowledge of the steady or mean flow.

Dynamic mode decomposition (DMD) is a data-driven technique widely used to analyze and model fluid problems including transonic buffet flows. Despite its strengths, DMD is known to suffer from sensitivities to the selected settings and the characteristics of the employed data. In this work, we closely examine the aforementioned sensitivities, identify possible pitfalls, and provide best practices to robustly perform DMD on a flow exhibiting transonic shock buffet. Specifically, we assess various DMD variants and test their sensitivity to the POD rank truncation and the sampling rate. A critical enabler to our analysis is a new presentation of the DMD algorithm as a modular framework consisting of five distinct steps. The tests also highlight the existing dangers of aliasing, when the sampling rate is too low. Finally, a list of practical recommendations and guidelines on how to accurately and robustly perform DMD on a transonic buffet flow is provided.

This work is concerned with uncertainty quantification in reduced-order dynamical system identification. Reduced-order models for system dynamics are ubiquitous in design and control applications and recent efforts focus on their data-driven construction. Our starting point is the sparse-identification of nonlinear dynamics (SINDy) framework, which reformulates system identification as a regression problem, where unknown functions are approximated from a sparse subset of an underlying library. In this manuscript, we formulate this system identification method in a Bayesian framework to handle parameter and structural model uncertainties. We present a general approach to enforce sparsity, which builds on the recently introduced class of neuronized priors. We perform comparisons between different variants such as Lasso, horseshoe, and spike and slab priors, which are all obtained by modifying a single activation function. We also outline how state observation noise can be incorporated with a probabilistic state-space model. The resulting Bayesian regression framework is robust and simple to implement. We apply the method to two generic numerical applications, the pendulum and the Lorenz system, and one aerodynamic application using experimental measurements., Comment: 15 pages, 10 figures

We propose a universal method for data-driven modeling of complex nonlinear dynamics from time-resolved snapshot data without prior knowledge. Complex nonlinear dynamics govern many fields of science and engineering. Data-driven dynamic modeling often assumes a low-dimensional subspace or manifold for the state. We liberate ourselves from this assumption by proposing cluster-based network modeling (CNM) bridging machine learning, network science, and statistical physics. CNM only assumes smoothness of the dynamics in the state space, robustly describes short- and long-term behavior and is fully automatable as it does not rely on application-specific knowledge. CNM is demonstrated for the Lorenz attractor, ECG heartbeat signals, Kolmogorov flow, and a high-dimensional actuated turbulent boundary layer. Even the notoriously difficult modeling benchmark of rare events in the Kolmogorov flow is solved. This automatable universal data-driven representation of complex nonlinear dynamics complements and expands network connectivity science and promises new fast-track avenues to understand, estimate, predict and control complex systems in all scientific fields.

Resolvent analysis identifies the most responsive forcings and most receptive states of a dynamical system, in an input--output sense, based on its governing equations. Interest in the method has continued to grow during the past decade due to its potential to reveal structures in turbulent flows, to guide sensor/actuator placement, and for flow control applications. However, resolvent analysis requires access to high-fidelity numerical solvers to produce the linearized dynamics operator. In this work, we develop a purely data-driven algorithm to perform resolvent analysis to obtain the leading forcing and response modes, without recourse to the governing equations, but instead based on snapshots of the transient evolution of linearly stable flows. The formulation of our method follows from two established facts: $1)$ dynamic mode decomposition can approximate eigenvalues and eigenvectors of the underlying operator governing the evolution of a system from measurement data, and $2)$ a projection of the resolvent operator onto an invariant subspace can be built from this learned eigendecomposition. We demonstrate the method on numerical data of the linearized complex Ginzburg--Landau equation and of three-dimensional transitional channel flow, and discuss data requirements. The ability to perform resolvent analysis in a completely equation-free and adjoint-free manner will play a significant role in lowering the barrier of entry to resolvent research and applications.

Periodically forced, oscillatory fluid flows have been the focus of intense research for decades due to their richness as a nonlinear dynamical system and their relevance to applications in transportation, aeronautics, and energy conversion. Recently, it has been observed that turbulent bluff-body wakes exhibit a subharmonic resonant response when excited with specific spatial symmetries at twice the natural vortex shedding frequency, which is hypothesized to be caused by triadic interactions. The focus of this paper is to provide new physical insight into the dynamics of turbulent oscillator flows, based on improved mechanistic models informed by a comprehensive experimental study of the turbulent wake behind a D-shaped body under periodic forcing. We confirm for the first time the role of resonant triadic interactions in the forced flow by studying the dominant components in the power spectra across multiple excitation frequencies and amplitudes. We then develop an extended Stuart-Landau model for the forced global wake mode, incorporating parametric and non-harmonic forcing. This model captures the system dynamics and reveals the boundaries of multiple synchronization regions. Further, it is possible to identify model coefficients from sparse measurement data, making it applicable to a wide range of turbulent oscillator flows. We believe these generalized synchronization models will be valuable for prediction, control, and understanding of the underlying physics in this ubiquitous class of flows., Comment: 10 pages, 5 figures

We propose an automatable data-driven methodology for robust nonlinear reduced-order modelling from time-resolved snapshot data. In the kinematical coarse-graining, the snapshots are clustered into few centroids representable for the whole ensemble. The dynamics is conceptualized as a directed network, where the centroids represent nodes and the directed edges denote possible finite-time transitions. The transition probabilities and times are inferred from the snapshot data. The resulting cluster-based network model constitutes a deterministic-stochastic grey-box model resolving the coherent-structure evolution. This model is motivated by limit-cycle dynamics, illustrated for the chaotic Lorenz attractor and successfully demonstrated for the laminar two-dimensional mixing layer featuring Kelvin-Helmholtz vortices and vortex pairing, and for an actuated turbulent boundary layer with complex dynamics. Cluster-based network modelling opens a promising new avenue with unique advantages over other model-order reductions based on clustering or proper orthogonal decomposition.

This manuscript describes the software package SCOUT, which analyzes, characterizes, and corrects one-dimensional signals. Specifically, it allows to check and correct for stationarity, detect spurious samples, check for normality, check for periodicity, filter, perform spectral analysis, determine the integral time scale, and perform uncertainty analysis on individual and on propagated signals through a data reduction equation. The novelty of SCOUT lies in combining these various methods into one compact and easy-to-use toolbox, which enables students and professionals alike to analyze, characterize, and correct for signals without expert knowledge. The program is oriented towards time traces, but an easy adaptation to spatial distributions can be performed by the user. SCOUT is available in two variants: a graphical user interface (GUI) and a script-based version. A key motivation of having two variants is to offer maximum flexibility to adaptively and visually adjust the analysis settings using the GUI version and to enable large batch processing capabilities and own code-integration using the script-based version. The package includes both variants as well as three example scripts with their corresponding signals.

We compute, model, and predict drag reduction of an actuated turbulent boundary layer at a momentum thickness based Reynolds number of Re{\theta} = 1000. The actuation is performed using spanwise traveling transversal surface waves parameterized by wavelength, amplitude, and period. The drag reduction for the set of actuation parameters is modeled using 71 large-eddy simulations (LES). This drag model allows to extrapolate outside the actuation domain for larger wavelengths and amplitudes. The modeling novelty is based on combining support vector regression for interpolation, a parameterized ridgeline leading out of the data domain, scaling from Tomiyama and Fukagata (2013), and a discovered self-similar structure of the actuation effect. The model yields high prediction accuracy outside the training data range.

Wall-resolved large-eddy simulations are performed to study the impact of spanwise traveling transversal surface waves in zero-pressure gradient turbulent boundary layer flow. Eighty variations of wavelength, period, and amplitude of the space- and time-dependent sinusoidal wall motion are considered for a boundary layer at a momentum thickness based Reynolds number of $Re_\theta = 1000$. The results show a strong decrease of friction drag of up to $26\,\%$ and considerable net power saving of up to $10\,\%$. However, the highest net power saving does not occur at the maximum drag reduction. The drag reduction is modeled as a function of the actuation parameters by support vector regression using the LES data. A substantial attenuation of the near-wall turbulence intensity and especially a weakening of the near-wall velocity streaks are observed. Similarities between the current actuation technique and the method of a spanwise oscillating wall without any normal surface deflection are reported. In particular, the generation of a directional spanwise oscillating Stokes layer is found to be related to skin-friction reduction.

We present the first general metric for attractor overlap (MAO) facilitating an unsupervised comparison of flow data sets. The starting point is two or more attractors, i.e., ensembles of states representing different operating conditions. The proposed metric generalizes the standard Hilbert-space distance between two snapshots to snapshot ensembles of two attractors. A reduced-order analysis for big data and many attractors is enabled by coarse-graining the snapshots into representative clusters with corresponding centroids and population probabilities. For a large number of attractors, MAO is augmented by proximity maps for the snapshots, the centroids, and the attractors, giving scientifically interpretable visual access to the closeness of the states. The coherent structures belonging to the overlap and disjoint states between these attractors are distilled by few representative centroids. We employ MAO for two quite different actuated flow configurations: (1) a two-dimensional wake of the fluidic pinball with vortices in a narrow frequency range and (2) three-dimensional wall turbulence with broadband frequency spectrum manipulated by spanwise traveling transversal surface waves. MAO compares and classifies these actuated flows in agreement with physical intuition. For instance, the first feature coordinate of the attractor proximity map correlates with drag for the fluidic pinball and for the turbulent boundary layer. MAO has a large spectrum of potential applications ranging from a quantitative comparison between numerical simulations and experimental particle-image velocimetry data to the analysis of simulations representing a myriad of different operating conditions., Comment: 33 pages, 20 figures

Shape optimization of an active and a passive drag-reducing device on a two-dimensional D-shaped bluff body is performed. The two devices are: Coanda actuator, and randomly-shaped trailing-edge flap. The optimization sequence is performed by coupling the genetic algorithm software DAKOTA to the mesh generator Pointwise and to the CFD solver OpenFOAM. For the the active device the cost functional is the power ratio, whereas for the passive device it is the drag coefficient. The optimization leads to total power savings of $\approx 70\%$ for the optimal Coanda actuator, and a 40\% drag reduction for the optimal flap. This reduction is mainly achieved through streamlining the base flow and suppressing the vortex shedding. The addition of either an active or a passive device creates two additional smaller recirculation regions in the base cavity that shifts the larger recirculation region away from the body and increases the base pressure. The results are validated against more refined URANS simulations for selected cases.

A new method for optimal sensor placement based on variable importance of machine learned models is proposed. With its simplicity, adaptivity, and low computational cost, the method offers many advantages over existing approaches. The new method is implemented on the flow over an airfoil equipped with a Coanda actuator. The analysis is based on flow field data obtained from 2D unsteady Reynolds averaged Navier-Stokes (URANS) simulations with different actuation conditions. The optimal sensor locations is compared against the current de-facto standard of maximum POD modal amplitude location, and against a brute force approach that scans all possible sensor combinations. The results show that both the flow conditions and the type of sensor have an effect on the optimal sensor placement, whereas the choice of the response function appears to have limited influence.