Machine learning flow control with few sensor feedback and measurement noise

TL;DR

This paper compares machine learning methods for active flow control around a cylinder, demonstrating how DRL and LGPC can effectively reduce drag with different robustness and interpretability features.

Contribution

It provides a comparative assessment of DRL and LGPC for flow control, highlighting their strengths and potential for combined approaches.

Findings

Both methods successfully stabilize the vortex wake and reduce drag.

DRL shows higher robustness to noise and initial condition variations.

LGPC produces interpretable control laws with fewer sensors.

Abstract

A comparative assessment of machine learning (ML) methods for active flow control is performed. The chosen benchmark problem is the drag reduction of a two-dimensional K\'arm\'an vortex street past a circular cylinder at a low Reynolds number ($Re=100$). The flow is manipulated with two blowing/suction actuators on the upper and lower side of a cylinder. The feedback employs several velocity sensors. Two probe configurations are evaluated: 5 and 11 velocity probes located at different points around the cylinder and in the wake. The control laws are optimized with Deep Reinforcement Learning (DRL) and Linear Genetic Programming Control (LGPC). By interacting with the unsteady wake, both methods successfully stabilize the vortex alley and effectively reduce drag while using small mass flow rates for the actuation. DRL has shown higher robustness with respect to variable initial conditions and to noise contamination of the sensor data; on the other hand, LGPC is able to identify compact and interpretable control laws, which only use a subset of sensors, thus allowing reducing the system complexity with reasonably good results. Our study points at directions of future machine learning control combining desirable features of different approaches.