Session G17: Flow Control: Optimization

Finding an optimal flow control with multi-point penalty method

Finding an optimal control of turbulent flows is inherently difficult due to its chaotic dynamics. The objective functional becomes extremely non-convex, having many local extrema that impede search strategies for finding effective controls. A multi-step penalty method, which temporarily introduces discontinuities at intermediate time points before minimizing them along the optimization, is demonstrating to be a promising approach to this challenge. However, while it is shown to find an effective control for turbulence, it can be still challenging for the flows in free space such as jets, where the controllable subspace is confined to a spatially limited region, near the actuator. We further develop additional techniques to extend the multi-step method to such flows in a spatially extensive domain.  A compact support for intermediate state is introduced, where the multi-step method can be selectively applied only to the region that exhibits chaos. A modified L₂-norm is suggested as a better penalty for flow optimization problem. This approach is demonstrated in a model system then demonstrated on a two-dimensional shear-layer noise-control application.   

Taming shear flows with gradient-enriched machine learning control

We propose an automated gradient-enriched machine learning control (gMLC, [1]) for fast learning of multiple-input multiple-output feedback control laws directly from the plant. gMLC is demonstrated on three shear flows: A DNS of the flow past a cluster of three rotating cylinders—the fluidic pinball—an open cavity flow experiment, and a smart skin separation control experiment. The fluidic pinball has been stabilized reducing the residual fluctuation energy up to 80%. For the open cavity, a mode-switching regime is fully stabilized with low actuation power [2]. For both cases, the need of feedback for the stabilization has been demonstrated. gMLC even learned smart skin separation control with 60 actuation commands and 54 sensors in few hours testing time (more: S. Li in this session). Key enablers are automated machine learning algorithms augmented with intermediate gradient steps: explorative gradient method for parametric optimization and gMLC for feedback law optimization. gMLC learns control laws significantly faster than previously employed feedback control strategies.   

Smart Skin Separation Control Exemplified for a Smooth Ramp

We perform the first AI-based smart skin experiment which utilizes distributed inputs and outputs (DIDO) to minimize flow separation over a smooth ramp. In this plant, flow separation may happen over a large range of locations. With smart skin deployed on the ramp surface, the flow separation can be progressively delayed via distributed actuation and sensing. We utilize multimodal actuators to delay flow separation. This novel actuator is composed of a height-adjustable vortex generator and an embedded mini-jet actuator. We massively deploy multimodal actuators (currently 30) on the ramp surface to delay flow separation that may happen at any point on the surface. Flow sensing is accomplished via distributed measurements of wall pressure (currently 56) from an array of high accuracy pressure sensors. With this setup, the smart skin is capable to achieve passive, active, and combined flow control strategies. The control efficiency can be optimized via artificial intelligence algorithms. In this work, we optimize the closed-loop active control for the smart skin. The control algorithm adopts the Gradient-enriched Machine Learning Control (gMLC) method proposed in [1]. This method combines exploration and exploitation, enabling a fast optimization in the high-dimensional space. Within 1,000 training periods, the optimized control strategy can effectively increase the pressure recovery on the ramp surface. Measurements of the controlled flow confirm significant reduction of the flow separation. These results will guide future development of flow control experiments with distributed sensing and actuation.

    

Neural network controllers applied to flow control

We report the application of a novel control approach based on the backpropagation of neural network models of dynamical systems. By leveraging sampled open-loop data, we train black box models with control inputs capable of learning important features from nonlinear systems. A neural network controller (NNC) is trained as a control law in a recurrent approach through backpropagation in closed loop. The methodology is first applied to four low-dimensional nonlinear plants presenting different features such as chaos and limit cycles around different equilibrium types. We also apply NNC to the high-order Kuramoto-Sivashinsky equation so as to attenuate the propagation of convective instabilities. Finally, we apply the technique to a cylinder flow with the goal of reducing the effects of instabilities. Results suggest that NNC presents implementation advantages over gradient based model predictive control due to its lower evaluation cost.   

Reduction of noise in cold and hot supersonic jets using active flow control guided by a genetic algorithm

This study demonstrates an experimental platform for active jet noise reduction comprising of an automated system that performs a search for the optimal actuator locations and parameters, powered by a genetic optimization algorithm. Sideline noise reduction levels of 7.3 dB for a cold overexpanded (NPR=2.8) jet, beyond the state-of-the-art for jet noise reduction with air injection. The reduction in noise was achieved at a mass flow ratio of 1.4% of the main jet, requiring no prior knowledge of the flow physics to inform the placement of the actuators. The same actuator pattern was tested in hot conditions (NTR=1.88), achieving 4.7 dB sideline noise reduction.

Detailed examination of the solutions obtained unveils some of the mechanisms leveraged by the GA to accomplish these high levels of noise reduction through microphone measurements and Schlieren flow visualization. The GA found that actuating at the diverging wall of the convergent-divergent nozzle, where flow is supersonic, is very effective when combined with air injection near the nozzle lip, outside of the nozzle. Although the external actuation is effective at eliminating the screech tone, it is by actuating inside the nozzle at the diverging wall that sufficient disruption of the shock cell train can be achieved in order to reduce the broadband shock-associated noise.