Machine learning and active flow control are expected to be the
next paradigm shift in fluid mechanics. They are critical enablers to lift increase, drag reduction, mixing enhancement, and noise reduction. The workgroup Flow Modelling and Control research
strategy is built around the closed-loop control system (assuming open-loop is a special case thereof). As illustrated in the figure below, the research can be classified into 4
We define a plant here as an experiment or
a model/simulation of an aerodynamic problem. A range fluid and aerodynamic problems are regularly investigated:
- Experiments: Experimental investigations in the various wind or water tunnel include:
- Drag reduction
experiments to increase efficiency and to lower fuel consumption. This concerns both ground- as well as air-based vehicles. The research on both ground- and air-based
vehicles range from fundamental to very-applied investigations. Investigating fundamental problems is not only beneficial to test basic concepts and to understand the main flow physics, but it is
also a good stepping stone for students towards more complicated problems. Drag reductions are achieved through passive as well as active means.
- Lift control experiments are mainly geared towards aircraft research. Envisaged actuation is typically active. The focus is on take-off and landing configurations. Examples include
investigations on an airfoil with a fast active flap, and the collaborative research center CRC 880 that investigated Coanda actuation on a high-lift configuration.
- Research in Modeling
incorporates a wide range of activities.
- Response surface
models are used in a range of applications to study flow sensitivities, to infer the best actuation settings, and as surrogates for optimization. Recent successes in
response surface modeling with machine learning include accurate model construction using support vector regression and random forest algorithms.
- Dynamical models can shed light on the main flow physics and can also be used for closed-loop control. The pursued methodologies cover the entire
modeling spectrum from black-box to white-box models, as illustrated below:
- Purely algorithmic ‘black’
methods, such as support vector machines (SVM) and the cluster-based network models (CNM), usually yield high accuracy and are easy to construct. One very powerful advantage of such methods,
particularly CNM, is their ability to capture rare events that POD-based methods are inherently not capable of doing. Further development of CNM are already underway.
- Owing to their good accuracy and
interpretability, data-driven approaches, such as sparse identification of nonlinear dynamics (SINDY), dynamic mode decomposition (DMD), and genetic programming have recently gained a lot of
attention. All three methods are actively used in the Flow Modelling and Control group and are developed internally as well as in collaboration with external partners.
- POD-based Galerkin models are the
‘whitest’ of the employed models. They are directly obtained from the Navier-Stokes equations through Galerkin projection. Despite being physically-based, POD-Galerkin models typically require
additional treatment/calibration to render them more stable and/or more accurate. Current and future research on POD-Galerkin models are mainly focused on system identification with physical
- Most of the
above-mentioned modeling methods are either already integrated (POD-Galerkin, CNM) or soon-to-be incorporated (SINDY, DMD) into xROM, the in-house reduced-order modeling toolbox. The
modularity of xROM combined with the ability to read a broad range of data formats enables stronger collaboration and easier dissemination of the toolbox. xROM is continuously expanded and maintained
to include the latest algorithms.
- Aerodynamic MAchine Learning (AMAL) is a
personal project of Richard Semaan, which has been in development for the last two years. AMAL is a machine learning-based methodology to predict all relevant aerodynamic properties. The first
findings from AMAL and the corresponding software are expected to be published and released toward the end of 2020.
In this section, we bundle the sensor signal(s) for closed-loop control
with any signal output (e.g. PIV) that might be acquired during an experiment for post-processing or analysis purposes. The research opportunities are broad:
- The number of sensors and their placement is critical for accurate
model predictions and closed-loop control applications. Optimal sensor placement (OSP) reduces instrumentation cost and increases the accuracy of state estimators. Current research
efforts build on previously published work and include a project to determine the OSP over a vehicle for force and moment coefficient predictions. Research on OSP using variable ranking of machine
learning models shall be expanded to other applications with different constraints.
- Sparse spatial sampling (S3) is a recently-developed method that samples numerical meshes without any information loss. This algorithm addresses some problematic aspects of Big Data by massively
reducing the storage size, and thus enabling easier handling and analysis of the data. S3 is planned to be further developed within the research unit FOR 2895 “Unsteady flow and
interaction phenomena at High-Speed Stall conditions” to include compressible flows and other metrics.
- A good understanding of any flow dynamics demands accurate
identification and characterization of coherent structures and vortices. Previous work in the Flow Modelling and Control group using POD-filtered Q-criteria to detect and quantify
vortices is currently investigated to include other data-driven vortex identification methods.
- Data mining is
the process of examining large databases to generate new information and to discover patterns. These techniques are widely used in other fields and are slowly finding their ways into the fluid
mechanics' community. Existing research in the group includes k-means clustering and Metric for Attractor Overlap (MAO).
- Despite monumental progress during the last two decades, PIV challenges
still exist. These include real-time PIV and pressure from PIV, which can be addressed using recent progress in deep neural networks on optical flows.
The research activities in control are concentrated in machine learning
- Machine Learning Control (MLC) is a model-free, data-driven approach for closed-loop control, which aims to find an optimal control law that minimizes a pre-chosen cost functional. MLC has great advantages
over traditional methods: (i) it is model-free, (ii) it is powerful, since it can deliver a performance that is at worst on-par with the best open-loop results, and (iii) it is easy to
implement and to use. MLC research is continuously evolving to reduce training time and to increase robustness. Applications range from active flow control to robotics.
- So far MLC has only been experimentally deployed using Genetic
Programming (GP). MLC using GP has been successfully implemented on a D-shaped bluff body, and on a high-lift configuration in both the wind and water tunnel with up to 32 individually
controlled actuators. Current and near-future improvements include better pre-testing algorithm using clustering, and boosting robustness through constant optimization.
- Cluster-based control is
a natural extension to the research activities in clustering. Plans are already underway to test the concept of numerical as well as on experimental plants.
Active flow control (AFC) requires actuators to affect the flow in the
desired manner. The field of active flow control has witnessed tremendous growth in the variety of actuators. Research activities in the Flow Modelling and Control group include:
- Researching passive flaps, which is a clear oxymoron
in the context of AFC. However, passive actuation should not be ruled out for the purpose of being trendy. Passive actuation sometimes delivers aerodynamic gains that match those of active actuators
at no cost and with a much-reduced complexity. Future research builds on existing surrogate-based optimization methods to optimize passive flaps and to test them when active actuation is not
necessary or is not efficient.
- Coanda actuation has been the main actuation mechanism in the group’s portfolio during the last 7 years. Coanda actuation yields high authority over the flow, it is mechanically simple, and is
easily scalable. Future efforts shall focus on optimizing and customizing the actuator’s geometry and settings (e.g. duty cycle, frequency) for particular applications (road vehicle versus aircraft
wing) and specific power sources (compressor versus blower).
- Hybrid Coanda is a new actuation mechanism that has been recently conceptualized and experimentally tested. Future research on the subject includes optimizing the actuation variables and testing
the actuation mechanism on various aerodynamic problems.