Scala, Alessandro (2025) Experimental application of Machine Learning algorithms to the active control of wake and separated flows via synthetic jets. [Tesi di dottorato]

	Documento PDF scala_alessandro_38_PARZIALE.pdf Visibile a [TBR] Amministratori dell'archivio Download (42MB) | Richiedi una copia
	Documento PDF scala_alessandro_38_COMPLETO.pdf Visibile a [TBR] Amministratori dell'archivio Download (76MB) | Richiedi una copia

Abstract

Flow control is increasingly crucial to modern engineering applications, ranging from aerodynamic drag reduction to the mitigation of flow separation and manipulation of complex vortical structures. Despite substantial progress, developing robust and adaptive control strategies for realistic, three-dimensional, and highly turbulent experimental environments remains a major challenge. Synthetic jets, widely used due to their compactness, low cost, and ability to excite multiple flow scales, are considered in this study. Notably, the influence of the actuation waveshape supplied to synthetic jet actuators remains largely unexplored, offering significant potential for innovation. This thesis investigates the impact of non-conventional, non-parametric actuation waveshapes on synthetic jet performance, employing Machine Learning (ML) algorithms as functional optimizers to autonomously identify effective control signals for specific flow configurations. The proposed ML-based control framework is experimentally validated on two canonical problems: the wake behind a circular cylinder and wingtip vortices generated by a finite wing, through three dedicated experimental campaigns encompassing open-loop and closed-loop control of the cylinder wake and open-loop control of wingtip vortices. The controlled flow fields are characterized via PIV and S-PIV, complemented by modal analysis (POD, E-POD) and data-driven clustering (Multi-Dimensional Scaling and k-means clustering) to elucidate underlying control mechanisms. Results show that optimal open-loop actuation reduces cylinder drag by up to 9.8%, while closed-loop strategies, enhanced by real-time sensing, achieve up to 14% drag reduction with markedly lower actuation power, demonstrating the efficiency of feedback-based control. For wingtip vortices, the optimal ML-generated policy reduces mean streamwise vorticity by approximately 65%, outperforming traditional sinusoidal forcing and disrupting vortex coherence. Overall, this work demonstrates the feasibility and effectiveness of ML-driven flow control in experimental environments, providing new insights into actuation dynamics and control landscapes. The original contributions include: (i) the first experimental validation of closed-loop, ML-based synthetic jet control; (ii) evidence that hybrid sensor-autonomous control laws enhance drag-reduction performance; (iii) a detailed flow-physics analysis via velocimetry and data-driven methods; and (iv) the extension of ML-based control to complex and three-dimensional turbulent flows.

Downloads

Actions (login required)

Modifica documento