Bulletin of the American Physical Society
74th Annual Meeting of the APS Division of Fluid Dynamics
Volume 66, Number 17
Sunday–Tuesday, November 21–23, 2021; Phoenix Convention Center, Phoenix, Arizona
Session A19: Flow Control I |
Hide Abstracts |
Chair: Aniketh Kalur, University of Minnesota Room: North 132 ABC |
Sunday, November 21, 2021 8:00AM - 8:13AM |
A19.00001: Using ROMs and adjoint-ROMs to optimize the flow past pitching-plunging wings Bolun Xu, Mingjun Wei, John T Hrynuk When an arbitrary new motion trajectory is explored to improve aerodynamic performance of the flow past pitching-plunging wings, the large control space makes most optimization efforts too expensive to be feasible. In this work, an adjoint-based method is used to study and optimize the flow past a pitching-plunging NACA0012 airfoil. The advantage of an adjoint-based method is to keep the computational cost independent of the number of control parameters. However, adjoint-based optimization using direct numerical simulation (DNS) may still be expensive because the optimization process requires many iterations with DNS and adjoint-DNS computations. To further improve the optimization efficiency for potential real-time applications, this work developed approaches to build and integrate reduced-order models (ROM) and adjoint-ROMs in an optimization scheme. Three control parameters are studied: plunging amplitude, pitching amplitude, and the phase delay between pitching and plunging motions. It is shown that the adjoint approach using ROMs can drastically reduce the optimization cost while reaching the same optimal solution achieved by the adjoint approach using DNS. |
Sunday, November 21, 2021 8:13AM - 8:26AM |
A19.00002: Feedback Stabilization of Incompressible Flows using the Quadratic Constraint Framework Talha Mushtaq, Peter Seiler, Maziar S Hemati Flow instabilities can create high shear regions that are detrimental to the performance of engineering systems. In this talk, we propose a framework to synthesize globally stabilizing feedback controllers to suppress such instabilities. The framework exploits the fact that the nonlinearity in the incompressible Navier-Stokes equations (NSE) is energy conserving. This property is expressed as a quadratic constraint and juxtaposed with the linear dynamics of NSE for controller synthesis using semi-definite programming (SDP). The resulting framework can be used to design full-state and static-output feedback controllers that stabilize the flow regardless of the perturbation magnitude. We demonstrate the approach on a reduced-order model of plane Couette flow with body force actuation. |
Sunday, November 21, 2021 8:26AM - 8:39AM |
A19.00003: Data-Driven Modeling and Control of Oscillatory Instabilities in a Kolmogorov-like Flow Nicholas Conlin, Maziar S Hemati, Jeffrey R Tithof
|
Sunday, November 21, 2021 8:39AM - 8:52AM |
A19.00004: Machine Learning for "Self-Healing" Flow Control Alejandro Stefan-Zavala, Chris Dougherty, Morteza Gharib Fan array wind tunnels (fan arrays) are a novel wind tunnel design consisting of multiple, individually controlled fans. This modularity is well suited to generate complex flows, such as those faced by unmanned aerial vehicles flying at low altitudes. In our study of fan array control, we postulate the problem of maintaining a uniform flow profile downstream despite the failure of particular fan units, which is expected in practice with some fan arrays numbering in the thousands of fans. Thus making the flow 'self-healing.' For this, we explore the application of reinforcement learning on a system consisting of a fan array with "dead" fans blowing into a grid of pressure probes downstream. |
Sunday, November 21, 2021 8:52AM - 9:05AM |
A19.00005: Physics-informed model-based deep reinforcement learning for dynamic flow control Jian-Xun Wang, Xinyang Liu, Han Gao Model-based reinforcement learning (MBRL) is believed to have much higher sample efficiency compared to model-free algorithms by learning a predictive model of the environment. However, the performance of MBRL highly relies on the quality of the learned model, which is usually built in a black-box manner and may have poor predictive accuracy outside of the data distribution. The deficiencies of the learned model may prevent the policy from being fully optimized. In this work, we developed a physics-informed MBRL framework, where governing equations and physical constraints are utilized to inform the model learning and policy search. By incorporating the prior information of the environment, the quality of the learned model can be notably improved, while the required interactions with the environment are significantly reduced, leading to better sample efficiency and learning performance. Moreover, to effectively capture the long-span transition dynamics of fluid flow with irregular domains, a novel network architecture based on graph embedding and attention-based transformer is developed, and the effectiveness has been demonstrated on both incompressible and compressible flows. |
Sunday, November 21, 2021 9:05AM - 9:18AM |
A19.00006: Deep Reinforcement Learning Using Data-Driven Reduced-Order Models Discovers and Stabilizes Low Dissipation Equilibria Kevin Zeng, Alec Linot, Michael D Graham Deep reinforcement learning (RL) is a data-driven method capable of discovering complex control strategies for high-dimensional systems, making it promising for flow control. However, a major challenge of RL is that substantial training data must be generated by interacting with the target system, making it costly when the flow system is computationally or experimentally expensive. We mitigate this challenge in a data-driven manner by combining dimension reduction via an autoencoder with a neural ODE framework to obtain a low-dimensional dynamical model from just a limited data set. We substitute this ROM in place of the true system during RL training to efficiently estimate the optimal policy, which is then deployed to the true system. We apply our method to the Kuramoto-Sivashinsky equation (KSE), a proxy system for turbulence that displays spatiotemporal chaos, equipped with equidistant actuators and demonstrate that we can learn a sufficient ROM of the actuated dynamics. With this ROM and a goal of minimizing dissipation and power cost, we extract control policies from it using RL. We show that the ROM-based strategies translate well to the KSE and highlight that the RL agent discovers and stabilizes a forced equilibrium solution. |
Sunday, November 21, 2021 9:18AM - 9:31AM |
A19.00007: Deep Reinforcement Learning for Active Drag Reduction in Wall Turbulence Luca Guastoni, Ali Ghadirzadeh, Jean Rabault, Philipp Schlatter, Hossein Azizpour, Ricardo Vinuesa Reinforcement Learning is a framework in which an agent learns to take decisions through a trial-and-error process. When the learning process is supported by deep neural networks, it is referred to as deep reinforcement learning (DRL). |
Sunday, November 21, 2021 9:31AM - 9:44AM |
A19.00008: A time parallelised adjoint-based optimization strategy applied to incompressible flow configurations Serena Costanzo, Taraneh Sayadi, Miguel Fosas de Pando, Peter J Schmid, Pascal Frey Adjoint-based methods are widely used in various areas of fluid mechanics as a cost effective way of evaluating gradient information in order to perform sensitivity analysis, control, data assimilation, etc. When applied to unsteady configurations, however, the calculation of the direct-adjoint loop requires the use of checkpointing algorithms, causing the optimisation procedure to become time consuming and in some cases even infeasible. One common approach to overcome this problem is using parallelisation. While parallelisation in space has been widely used to reduce the cost of CFD calculations, time parallelisation is less explored. This is to a great extend due to unpredictable convergence rate of existing time parallel methods when applied to highly nonlinear and unsteady complex flow regimes. The linear nature of the adjoint equations, however, make them suitable for such implementations. In this study, we introduce a parallel in time algorithm designed to speed up the integration of the adjoint equations and ultimately the optimisation procedure. The code used to perform the calculations is a two-dimensional incompressible Navier-Stokes solver with immersed boundaries capabilities. Various control strategies are considered from drag reduction around a shedding cylinder to reducing pressure loss across a blade. In all cases, the parallel in time strategy allows the extraction of the gradient at a fraction of a time compared to the time used in the direct adjoint loop. |
Sunday, November 21, 2021 9:44AM - 9:57AM |
A19.00009: Stochastic optimisation of the flow around a linear cascade of blades Alejandro Quirós Rodríguez, Taraneh Sayadi, Miguel Fosas de Pando The past decades have seen remarkable progress in computing capabilities, allowing computational fluid dynamics (CFD) to become an ever more present tool in describing and predicting complex unsteady flows. However, robust optimisation and control of these flows on the basis of such high-fidelity simulations remains a big challenge. The main bottleneck arises from the large cost associated with performing each function evaluation (a full and potentially unsteady CFD calculation). As the first step, we study the performance of a stochastic optimisation algorithm using response surfaces (DYCORS) when applied to such cases. The performance of this method is then improved by adding local gradient information as well as the functional value in constructing the response surface. The effectiveness of the proposed algorithm is then studied by optimising the total pressure drop around a linear cascade of blades, by imposing a tangential velocity in the blade surface. A range of Reynolds numbers are considered and the performance of the method is compared to the original derivative free algorithm as well as a gradient-based alternative. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700