Artificial intelligence control applied to drag reduction of the fluidic pinball
Résumé
Flow control is at the core of many engineering applications, such as drag reduction for land-, sea- and air-borne tranportation, lift increase of airfoils, mixing enhancement of chemical reactions and efficiency increase for wind turbines and marine current power, just to name a few examples. In particular, feedback turbulence control is a key enabler for those technologies (Brunton & Noack, 2015 Appl. Mech. Rev. 67, 050801).
The challenges in control design are the associated high-dimensional dynamics, significant time-delays from actuation to sensing, strong nonlinearities and frequency crosstalk, making modeling and linear control theory often impractical.
To address these challenges, a model-free, self-learning control strategy is developed building on corresponding experimental studies of the authors.
The control design is formulated as a regression problem and solved with powerful methods of machine learning / artificial intelligence .
Our machine learning control (MLC, Duriez et al. 2016 Springer) is based on genetic programming (GP), a biological inspired method that mimicks the Darwinian process of natural selection to learn the control law in an unsupervised way.
MLC is applied to a direct numerical simulation of the fluidic pinball which is taken as a drag reduction benchmark.
Here, three cylinders are centered on an equilateral triangle in a two dimensional flow.
The flow is actuated by three independent cylinder rotations and monitored by 9 sensors downstream.
The fluidic pinball is a simple configuration which shares nonlinear features of many reported actuated real flows such as bifurcations and frequency crosstalk.
The goal is to reduce the parasitic drag power accounting for the actuation expenditure.
On the methodological side, we increase the GP learning rate by a factor of 5, for instance, by parameter optimization and removing redundant control laws.
On the physical side, we achieve 46% reduction of the net drag power comprising open-loop boat-tailing and closed-loop phasor control after testing only 1000 individuals. This reduction outperforms all other optimized symmetric actuations (like constant and periodic forcing). We expect significant performance improvements of MLC in future turbulence control experiments.