Fluid Mechanics and Its Applications

Thomas Duriez
Steven L. Brunton
Bernd R. Noack

Machine Learning
Control –
Taming Nonlinear
Dynamics and
Turbulence
Fluid Mechanics and Its Applications

Volume 116

Series editor
André Thess, German Aerospace Center, Institute of Engineering
Thermodynamics, Stuttgart, Germany

Founding Editor
René Moreau, Ecole Nationale Supérieure d’Hydraulique de Grenoble,
Saint Martin d’Hères Cedex, France
Aims and Scope of the Series

The purpose of this series is to focus on subjects in which fluid mechanics plays a
fundamental role.
As well as the more traditional applications of aeronautics, hydraulics, heat and
mass transfer etc., books will be published dealing with topics which are currently
in a state of rapid development, such as turbulence, suspensions and multiphase
fluids, super and hypersonic flows and numerical modeling techniques.
It is a widely held view that it is the interdisciplinary subjects that will receive
intense scientific attention, bringing them to the forefront of technological
advancement. Fluids have the ability to transport matter and its properties as well
as to transmit force, therefore fluid mechanics is a subject that is particularly open to
cross fertilization with other sciences and disciplines of engineering. The subject of
fluid mechanics will be highly relevant in domains such as chemical, metallurgical,
biological and ecological engineering. This series is particularly open to such new
multidisciplinary domains.
The median level of presentation is the first year graduate student. Some texts are
monographs defining the current state of a field; others are accessible to final year
undergraduates; but essentially the emphasis is on readability and clarity.

More information about this series at http://www.springer.com/series/5980

Thomas Duriez Steven L. Brunton
•

Bernd R. Noack

Machine Learning
Control –
Taming Nonlinear
Dynamics and
Turbulence

123
Thomas Duriez Bernd R. Noack
Laboratorio de Fluido Dinámica Département Mécanique-Energétique
CONICET - Universidad de Buenos Aires LIMSI-CNRS, UPR 3251
Buenos Aires Orsay
Argentina France
and
Steven L. Brunton
Mechanical Engineering Department Institut für Strömungsmechanik
University of Washington Technische Universität Braunschweig
Seattle, WA Braunschweig
USA Germany

ISSN 0926-5112 ISSN 2215-0056 (electronic)

Fluid Mechanics and Its Applications
ISBN 978-3-319-40623-7 ISBN 978-3-319-40624-4 (eBook)
DOI 10.1007/978-3-319-40624-4

Library of Congress Control Number: 2016943421

© Springer International Publishing Switzerland 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG Switzerland
Preface

This book is an introduction to machine learning control (MLC), a surprisingly

simple model-free methodology to tame complex nonlinear systems. These systems
are assumed to be manipulated by a finite number of actuators (inputs) and mon-
itored by a finite number of sensors (outputs). The control logic is chosen to
minimize a well-defined cost functional.
MLC brings together three well-established disciplines: the theory of
closed-loop feedback control, machine learning and regression, and the nonlinear
dynamical systems that are characteristic of turbulent fluid flows. Over the past
decades, control theory has developed into a mature discipline with a beautiful
theoretical foundation and powerful associated numerical algorithms. Important
advances have been made to enable robust control of systems with sensor noise,
external disturbances, and model uncertainty. Modern methods from control theory
now pervade the engineering sciences and have transformed the industrial land-
scape. However, challenges remain for the control of systems with strongly non-
linear dynamics leading to broadband frequency spectra, a high-dimensional state
space, and large time delays. MLC begins to address these challenges using
advanced methods from machine learning to discover effective control laws.
Many turbulence control problems are not adequately described by linear
models, have exceedingly large state spaces, and suffer from time delays from
actuators to sensors via nonlinear convective fluid dynamic effects. Take for
instance the aerodynamic drag minimization of a car with actuators at the back side,
pressure sensors distributed over the car, and a smart feedback control logic.
Numerical simulation of the underlying dynamics given by the Navier–Stokes
equations requires days or weeks, while the control system requires actuation
decisions on the order of milliseconds. Reduced-order models that incorporate
nonlinearities, multiscale phenomena, and actuation effects have eluded many
serious efforts and will likely remain elusive for years to come. In short, there may
not even be a viable model for robust control design. Nevertheless, the literature
contains many studies on turbulence control, with the majority either employing
open-loop forcing such as periodic blowing, slowly adapting a working open-loop

vii
viii Preface

strategy, or stabilizing an underlying laminar solution such as a laminar boundary

layer on an aircraft wing. Feedback control responding to dominant flow structures
in real time may be found in numerical studies but is rarely found in real-world
experiments with turbulent flows. Indeed, turbulence control is a grand challenge
problem in engineering, with far-reaching potential scientific, industrial, and soci-
etal impact.
Yet, just looking at the flight maneuvers of a bird, bat, or insect, it is clear that
nature has found impressive feedback flow control solutions without employing
advanced mathematical models. An eagle, for instance, can land gently on a rocky
surface under gusty wind conditions and in rain by moving its wings and feathers to
expertly manipulate fluid forces. More than 50 years ago, Ingo Rechenberg and
Hans-Peter Schwefel have emulated nature’s evolutionary way to optimize flow
properties at the former Hermann-Föttinger-Institut of the Berlin Institute of
Technology, Germany. Their pioneering evolutionary strategies started evolution-
ary computations. Subsequent innovations include evolutionary programming
(Fogel Owens and Walsh 1966), genetic algorithms (Holland 1975) and genetic
programming (Koza 1992). These evolutionary computations constitute an
important pillar of machine learning. In a visionary publication in 1959, the arti-
ficial intelligence pioneer Arthur Samuel defined machine learning as a ‘field of
study that gives computers the ability to learn without being explicitly pro-
grammed.’ Machine learning is a rapidly evolving discipline of computer science
that is benefiting from the current explosion of big data. It has successfully
improved an immense range of technologies—from the smart phone to the autopilot
in Tesla’s Sedan and to large-scale factory processes. In academia, nearly all sci-
entific disciplines are profiting from machine learning.
Not surprisingly, machine learning methods may augment or replace control
design in myriad applications. Robots learn to walk with dynamic programming.
Genetic algorithms are used to optimize the coefficients in proportional–integral–
derivate (PID) controllers. The groups of N. Benard and E. Moreau employ genetic
algorithms to optimize linear sensor feedback in a flow control experiment.
Reinforcement learning has been successfully applied to stabilize chaotic dynamics,
and has recently also reduced cavity noise in an experiment operated by F.
Lusseyran, L. Pastur and L. Mathelin. The authors of this book have pushed the first
applications of genetic programming in feedback control of nonlinear dynamics,
direct Navier–Stokes simulations and experimental turbulent shear flows. This book
focuses on arguably one of the simplest, most versatile and yet very powerful
version of machine learning control: Optimal nonlinear control laws are identified
with genetic programming. Corresponding success stories are described throughout
this book.
The authors have taught material from this book in several university courses.
These courses focus on basic principles using simple examples, and the content
requires anywhere from 15 to 30 h to cover. Our students have had backgrounds in
computer science, control theory, nonlinear dynamics, or fluid mechanics. The
prospective reader is not expected to have hands-on expertise in any of these fields
but should come with the ambition to control a complex system. The book is
Preface ix

organized as follows. In Chap. 1, the reader is introduced to feedback control and its
challenges for complex real-world problems. Chapter 2 constitutes the core of the
book. This chapter formulates feedback control as a regression problem and
employs genetic programming as a powerful regression technique to identify the
best feedback law. Chapter 3 reviews classical methods of control theory against
which MLC is benchmarked in Chap. 4 for linear and weakly nonlinear dynamics.
These chapters provide context for feedback control, but they are not required to
implement the MLC methods in Chap. 2. The hurried reader may jump to Chap. 5
if she/he is interested in strongly nonlinear dynamics applications or to Chap. 6
if she/he is interested in experimental implementations of feedback flow control.
Chapter 7 distills good practices for real-world experiments that need to be taken
into account in any MLC implementation. In Chap. 8 we provide an outlook on
future methodological advances, which are expected to drastically amplify the
applicability and performance of MLC. In addition, we list a number of future MLC
applications with epic proportions.
We have profited tremendously from interactions with many colleagues on
machine learning control. First, we highly appreciate André Thess for his continual
encouragement to write a book about turbulence control for this Springer series. He
has nurtured the idea for years before we decided to write this book. We highly
appreciate the insightful and inspiring interviews with leading scholars of the field:
Shervin Bagheri, Belinda Batten, Mark Glauser, Marc Schoenauer, and David
Williams. These additions provide valuable perspectives for past progress and
future work. Eurika Kaiser has provided continual exquisite feedback on our
chapters and also contributed with her illuminating visualizations in Chap. 7
showing the performance of MLC.
We have also benefited greatly from our mentors throughout our careers. BRN is
deeply indebted to his turbulence control mentors Andrzej Banaszuk, Andreas
Dillmann, Helmut Eckelmann, Rudibert King, and William K. George, who shared
and fueled the passion for the field. SLB would like to gratefully acknowledge and
thank Nathan Kutz, Naomi Leonard, Richard Murray, Clancy Rowley, and Rob
Stengel, who each found unique ways to make dynamics and control theory come
to life. TD would like to acknowledge Eduardo Jose Wesfreid, Jean-Luc Aider,
Guillermo Artana, Luc Pastur, François Lusseyran, and Bernd R. Noack, who each
have had a profound (and most beneficial) impact on his perception of the different
fields he has been in contact with. This book would not have been possible without
our many colleagues, collaborators, and co-authors who have shared our early
enthusiasm for MLC and have dedicated significant energy to developing it:
Markus Abel, Jean-Luc Aider, Zhe Bai, Diogo Barros, Jean-Paul Bonnet, Jacques
Borée, Bing Brunton, Juan Martin Cabaleiro, Camila Chevot, Tom Daniel, Antoine
Debien, Laurent Cordier, Christophe Cuvier, Joël Delville (d), Caroline Fourment
(d), Hiroaki Fukumoto, Nicolas Gautier, Fabien Harambat, Eurika Kaiser, Laurent
Keirsbulck, Azeddine Kourta, Kai von Krbek, Nathan Kutz, Jean-Charles
Laurentie, Ruiying (Cecile) Li, François Lusseyran, Robert Martinuzzi, Lionel
Mathelin, Nicolas Mazellier, Marek Morzyński, Christian Nayeri, Robert Niven,
Akira Oyama, Vladimir Parezanović, Oliver Paschereit, Luc Pastur, Brian Polagye,
x Preface

Josh Proctor, Bartosz Protas, Rolf Radespiel, Cedric Raibaudo, Jim Riley,
Tony Ruiz, Michael Schlegel, Peter Scholz, Marc Segond, Richard Semaan, Tamir
Shaqarin, Andreas Spohn, Michel Stanislas, Ben Strom, and Sam Taira. Many of
our co-authors have applied the nascent MLC methodology in their own experi-
ments early on, when success was far from certain. We thank our students for
visiting our courses in Argentina, France, Germany, and the USA and contributing
with many good questions, new ideas and encouraging project results. Anneke Pot
from Springer Publisher has dependably supported us in critical decisions about
book contents and the production procedure.

Buenos Aires Thomas Duriez

Seattle Steven L. Brunton
Paris-Saclay Bernd R. Noack
April 2016
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Feedback in Engineering and Living Systems . . . . . . . . . . . . . . . 1
1.2 Benefits of Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Challenges of Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Feedback Turbulence Control is a Grand Challenge Problem . . . . 7
1.5 Nature Teaches Us the Control Design. . . . . . . . . . . . . . . . . . . . 8
1.6 Outline of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Machine Learning Control (MLC) . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Methods of Machine Learning. . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 System Identification as Machine Learning . . . . . . . . . . . . 13
2.1.2 Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.3 Genetic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.4 Additional Machine Learning Methods . . . . . . . . . . . . . . 18
2.2 MLC with Genetic Programming. . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.1 Control Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.2 Parameterization of the Control Law . . . . . . . . . . . . . . . . 20
2.2.3 Genetic Programming as a Search Algorithm . . . . . . . . . . 21
2.2.4 Initializing a Generation . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.5 Evaluating a Generation . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.6 Selecting Individuals for Genetic Operations. . . . . . . . . . . 26
2.2.7 Selecting Genetic Operations . . . . . . . . . . . . . . . . . . . . . 27
2.2.8 Advancing Generations and Stopping Criteria . . . . . . . . . . 30
2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3.1 Fitting a Function Through Data Points . . . . . . . . . . . . . . 33
2.3.2 MLC Applied to Control a Dynamical System . . . . . . . . . 36
2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.5 Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.6 Interview with Professor Marc Schoenauer . . . . . . . . . . . . . . . . . 46

xi
xii Contents

3 Methods of Linear Control Theory. . . . . . . . . . . . . . . . . . . . . . . . . 49

3.1 Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Full-State Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3 Sensor-Based State Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4 Sensor-Based Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.5 System Identification and Model Reduction . . . . . . . . . . . . . . . . 58
3.5.1 System Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.2 Eigensystem Realization Algorithm (ERA) . . . . . . . . . . . . 59
3.5.3 Observer Kalman Filter Identification (OKID). . . . . . . . . . 62
3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.7 Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4 Benchmarking MLC Against Linear Control . . . . . . . . . . . . . .... 69
4.1 Comparison of MLC with LQR on a Linear Oscillator . . . . . .... 70
4.2 Comparison of MLC with Kalman Filter on a Noisy Linear
Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3 Comparison of MLC with LQG for Sensor-Based Feedback . . . . . 80
4.4 Modifications for Small Nonlinearity . . . . . . . . . . . . . . . . . . . . . 84
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.6 Interview with Professor Shervin Bagheri . . . . . . . . . . . . . . . . . . 89
5 Taming Nonlinear Dynamics with MLC . . . . . . . . . . . . . . . . . . . . . 93
5.1 Generalized Mean-Field System . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2 Machine Learning Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2.1 Formulation of the Control Problem . . . . . . . . . . . . . . . . 98
5.2.2 MLC Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2.3 MLC Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3 Derivation Outline for the Generalized Mean-Field Model . . . . . . 105
5.4 Alternative Control Approaches. . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4.1 Open-Loop Forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4.2 Closed-Loop Forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.4.3 Short-Term Forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.6 Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.7 Interview with Professor Mark N. Glauser . . . . . . . . . . . . . . . . . 117
6 Taming Real World Flow Control Experiments with MLC . . . . . . . 121
6.1 Separation Control Over a Backward-Facing Step . . . . . . . . . . . . 122
6.1.1 Flow Over a Backward-Facing Step . . . . . . . . . . . . . . . . 122
6.1.2 Experimental Setup at PMMH . . . . . . . . . . . . . . . . . . . . 123
6.1.3 Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2 Separation Control of Turbulent Boundary Layers . . . . . . . . . . . . 128
6.2.1 Separating Boundary Layers . . . . . . . . . . . . . . . . . . . . . . 128
6.2.2 Experimental Setups at LML and PRISME. . . . . . . . . . . . 129
6.2.3 Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Contents xiii

6.3 Control of Mixing Layer Growth. . . . . . . . . . . . . . . . . . . . . . . . 135

6.3.1 Mixing Layer Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.3.2 Experimental Setup of the TUCOROM Wind Tunnel . . . . 136
6.3.3 Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.4 Alternative Model-Based Control Approaches . . . . . . . . . . . . . . . 140
6.5 Implementation of MLC in Experiments. . . . . . . . . . . . . . . . . . . 143
6.5.1 Real-Time Control Loop—from Sensors to Actuators . . . . 143
6.5.2 MLC Implementation in the PMMH Flow Over
a Backward-Facing Step. . . . . . . . . . . . . . . . . . . . . . . . . 144
6.5.3 MLC Implementation in the LML and PRISME
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.5.4 MLC Implementation in the TUCOROM Experiment . . . . 146
6.6 Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.7 Interview with Professor David Williams . . . . . . . . . . . . . . . . . . 149
7 MLC Tactics and Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.1 The Ideal Flow Control Experiment . . . . . . . . . . . . . . . . . . . . . . 153
7.2 Desiderata of the Control Problem—From the Definition
to Hardware Choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.2.1 Cost Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.2.2 Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.2.3 Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.2.4 Search Space for Control Laws . . . . . . . . . . . . . . . . . . . . 157
7.3 Time Scales of MLC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.3.1 Controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.3.2 Response Time of the Plant . . . . . . . . . . . . . . . . . . . . . . 160
7.3.3 Learning Time for MLC . . . . . . . . . . . . . . . . . . . . . . . . 161
7.4 MLC Parameters and Convergence . . . . . . . . . . . . . . . . . . . . . . 162
7.4.1 Convergence Process and Its Diagnostics . . . . . . . . . . . . . 162
7.4.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.4.3 Pre-evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.5 The Imperfect Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.5.1 Noise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.5.2 Drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.5.3 Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
8 Future Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.1 Methodological Advances of MLC . . . . . . . . . . . . . . . . . . . . . . 170
8.2 System-Reduction Techniques for MLC—Coping with
High-Dimensional Input and Output. . . . . . . . . . . . . . . . . . . . . . 174
8.3 Future Applications of MLC . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
xiv Contents

8.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

8.5 Interview with Professor Belinda Batten . . . . . . . . . . . . . . . . . . . 184
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Matlab® Code: OpenMLC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Abbreviations

ANN Artificial Neural Network

ARMA(X) Auto-Regressive Moving Average (with eXogenous input)
AVERT Aerodynamic Validation of Emission Reducing Technologies
BPOD Balanced Proper Orthogonal Decomposition
CROM Cluster-based Reduced-Order Modeling
DEIM Discrete Empirical Interpolation Method
DMD Dynamic Mode Decomposition
EC Evolutionary Computing
EP Evolutionary Programming
ERA Eigensystem Realization Algorithm
GA Genetic Algorithm
GMFM Generalized Mean-Field Model
GP Genetic Programming
LML Laboratoire de Mécanique de Lille, Université de Lille 1 Cité
Scientifique, Bâtiment M3 - 59655 Villeneuve d’Ascq Cedex,
France
LPV Linear Parameter Varying
LQE Linear Quadratic Estimation
LQG Linear Quadratic Gaussian
LQR Linear Quadratic Regulator
MIMO Multiple Input Multiple Output
MISO Multiple Input Single Output
ML Machine Learning
MLC Machine Learning Control
NLSA Nonlinear Laplacian Spectral Analysis
OKID Observer Kalman filter IDentification
PCA Principal Component Analysis
PID Proportional–Integral–Derivative
PIV Particle Image Velocimetry

xv
xvi Abbreviations

PMMH Physique et Mecanique des Millieux Hétérogènes Laboratory, 10

rue Vauquelin, 75231 Paris Cedex, France
POD Proper Orthogonal Decomposition
PPRIME Institute Pôle Poitevin de Recherche pour l’Ingénieur en
Mécanique, Matériaux et Énergétique, 11 Boulevard Marie et Pierre
Curie BP 30179, 86962 Futuroscope Chasseneuil Cedex, France
PRISME Laboratoire Pluridisciplinaire de Recherche, Ingénierie des
Systémes, Mecanique, Energétique. Université d’Orléans 8 Rue
Léonard de Vinci, 45072 Orléans, France
ROM Reduced-Order Model
RT Real-Time
SIMO Single Input Multiple Output
SISO Single Input Single Output
SSA Singular Spectrum Analysis
SVM Support Vector Machine
TUCOROM TUrbulence COntrol using Reduced-Order Models, ANR Chair of
Excellence (ANR-10-CEXC-0015), Poitiers, France
UVG Unsteady Vortex Generator
Symbols

A; Ad ; A~ State matrix (continuous time; discrete time; reduced)

a; ak ; am ; a State (vector, continuous time; vector, discrete time kth step;
mth component; scalar)
^a; ^
ak Full-state estimate (continuous; discrete time)
B; Bd ; B~ Input matrix (continuous; discrete time; reduced)
B Amplitude of periodic forcing
b; bk ; bm ; b Actuation command (vector, continuous time; vector, discrete
time kth step; mth component; scalar)
B Matrix of control inputs
C; Cd ; C ~ Output matrix (continuous; discrete time; reduced)
cμ Momentum coefficient
C; Cd Controllability matrix (continuous; discrete time)
D; Dd ; D ~ Feedthrough matrix (continuous; discrete time, reduced)
D Characteristic distance
dc Duty cycle
ex , ey , ez Unity vectors associated with directions x, y and z
E Expectation operator
F Dynamics
FD Drag force
G Measurement function
g Gain of control command in a generalized mean-field model
H Hankel matrix
H Heavyside function
Hsection Height of the test section
hi ðtÞ; hi;u ; hi;max hot-wire or hot-film signal number i (raw signal average value
of the unactuated measurement average measurement under
constant maximal actuation)
hstep ; hramp Height (of the step of the ramp)
I Identity matrix
i Index of individual (or other counter)

xvii
xviii Symbols

J; Jij Cost function value; of individual i in generation j

Ja Cost on states
Jb Cost on actuation
j Index of generation
K Control function
Kf Kalman filter gain
Kr Regulator gain, full-state control matrix
L Length of the experimental test section
Lsep Separation length
l Width of the experimental test section
‘ Ramp length
Na Number of states
Nb Number of actuation commands
Ne Number of individuals concerned by elitism
Ng Number of generations
Ni Number of individuals
Np Tournament size
Ns Number of sensors
O; Od Observability matrix (continuous time; discrete time)
Pc Probability of crossover
Pm Probability of mutation
Pr Probability of replication
p Pressure
pðaÞ Probability density of states
Q State cost weight matrix for LQR
Q; Qu Flow rate to actuator jets (instantaneous; average value under
constant blowing)
R Actuation cost weight matrix for LQR
r , r Amplitude of oscillators of a generalized mean-field model
(Table 5.1)
Re Reynolds number
Sa ðtÞ; Sa;u Area of backflow (instantaneous; unactuated average value)
Sb Actuator cross section
Sj Jet cross section
Sref Ramp reference surface
s; sk ; sm ; s Sensor signal (vector, continuous time; vector, discrete time
kth step; mth component; scalar)
^s; ^
ak Expected sensor value (continuous time; discrete time)
S; S  Markov parameters; of the augmented system
T Evaluation time
Trms Time period used to compute RMS of hot-wire signal
fluctuations
t, t0 Time, initial time
U; Ur Left singular vectors of SVD (complete; reduced)
Symbols xix

U Characteristic velocity
u; us , uΔ , u , u Velocity (vector field; steady solution; deviation due to
Reynolds stresses; contribution of frequency ω ; contribution
of frequency ω )
u Slow varying mean flow
u
0
Flow fluctuations
u Streamwise velocity component
V; Vr Right singular vectors of SVD (complete reduced)
Vd Disturbance variance
Vn Noise variance
VJet Characteristic velocity of jets
v Velocity vector initial condition
v Transverse velocity component
WdC Discrete-time controllability Gramian
WdO Discrete-time observability Gramian
W Mixing layer width
w Disturbance array
wr External reference signal
wd External disturbance, process noise
wn Measurement noise
w Spanwise velocity component
X Solution to the Riccati equation for LQR
x Space vector
x Streamwise coordinate
Y Solution to the Riccati equation for Kalman filter
y Transverse coordinate
z System output
z Spanwise coordinate
β , β , β , β Parameter for growth rate change in oscillators of a general-
ized mean-field model (Table 5.1)
γ Penalization coefficient
γ  , γ  , γ  , γ  Parameter for frequency change in oscillators for a generalized
mean-field model (Table 5.1)
δðÞ Dirac delta function
ε Nonlinearity strength coefficient or state stabilization error
κ Gain of the generalized mean-field model
ν Kinematic viscosity
ρ
P P Fluid density
; r Singular values matrix of SVD, (complete; reduced)
σ Oscillator growth rate
σ  ; σ H ; σ  ; σ H Growth rate of oscillators of a generalized mean-field model
(Table 5.1)
τ; τa ; τu Period of time (with actuated system; with unactuated system)
xx Symbols

φ , φ Phase of oscillators in a generalized mean-field model

(Table 5.1)
χ Backflow coefficient
Ω Space domain
ω Oscillator pulsation
ω ; ωH ; ω ; ωH Frequency of oscillators in a generalized mean-field model
(Table 5.1)