Effect of Synthetic Jets Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder
summary
The gist
Optimization Synthetic Jet Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder investigates how the position and width of synthetic jets influence…
In short
This study used deep reinforcement learning to optimize synthetic jet parameters (position and width) for flow control around a square cylinder. The DRL agent successfully reduced drag and lift, suppressing vortex shedding. Optimal jet placement depends on the Reynolds number, with trailing edge jets best at lower Re and leading edge jets better at higher Re.
Key concepts
- Deep Reinforcement Learning (DRL)
- A machine learning approach where an 'agent' learns to control a system by interacting with it. The agent observes the flow state and takes actions (adjusting jet mass flow rates) to maximize a specific reward, effectively learning the best way to control the cylinder's flow autonomously.
- Synthetic Jets
- Small actuators placed on a surface that inject periodic jets of fluid into the surrounding flow. These jets are used here to actively manipulate the airflow around a square cylinder, helping to reduce drag and prevent vortices from forming in the wake.
- Drag Reduction/Vortex Suppression
- The goal of the control system is to minimize aerodynamic resistance (drag) and eliminate swirling eddies (vortices) behind the cylinder. The DRL agent learns how to adjust jet settings to achieve these performance improvements across different flow conditions.
Terminology used across episodes
This episode discusses
- Effect of Synthetic Jets Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder · Paper Radio
- Playing Atari with Deep Reinforcement Learning
- Continuous control with deep reinforcement learning
- Deep Reinforcement Learning that Matters
- Deep Reinforcement Learning: An Overview
- Model-Based Reinforcement Learning for Atari
- Deep reinforcement learning-based active flow control of an elliptical cylinder: transitioning from an elliptical cylinder to a circular cylinder and a flat plate · Paper Radio
- Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor
- Soft Actor-Critic Algorithms and Applications
- Reinforcement-learning-based control of confined cylinder wakes with stability analyses
- Applying deep reinforcement learning to active flow control in turbulent conditions
The paper
Effect of Synthetic Jets Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder · Read on arXiv
Wang Jia, Hang Xu
School of Ocean and Civil Engineering, Shanghai Jiao Tong University
DOI: 10.1063/5.0220149
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Effect of Synthetic Jets Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder".
Mira: Here's the gist of the paper. Optimization Synthetic Jet Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder investigates how the position and width of synthetic jets influence flow control performance using a deep reinforcement learning algorithm to achieve drag reduction and vortex suppression Core Findings on Flow Control Effectiveness The DRL agent, specifically a SAC agent, is found to be capable of achieving significant reductions in the lift and drag coefficients of a square cylinder, as well as completely suppress vortex shedding in the wake The main findings of the study are: 1.
Kai: DRL-based control is shown to effectively control flow around square cylinders at two different Reynolds numbers of 100 and 500 At Re = 100, synthetic jets located near the trailing edge corners achieve the maximum drag reduction effect, with the mean drag coefficient reduced by 14.4%, the standard deviation of the drag coefficient reduced by 99%, and the standard deviation of the lift coefficient reduced by 86.1% At Re = 500, synthetic jets located near the leading edge corners achieve a drag reduction rate of 65.5%, with the mean drag coefficient reduced by 51.4%, the standard deviation of the drag coefficient reduced by 90.5%, and the standard deviation of the lift coefficient reduced by 74.9% The DRL framework is developed to enable interaction between an agent and a numerical simulation environment, where the agent observes the current state of the CFD environment and adjusts the mass flow rates of two synthetic jets on the square cylinder based on reward feedback The specific operational process is depicted in Fig.
Mira: 1 The agent's actions are represented by dimensionless mass flow rates Q1 and Q2, with the condition Q1 +Q2 = 0 to maintain zero net mass flow The state st represents physical information collected from specific locations within the computational domain, specifically 201 probes are placed around the square cylinder and in the wake region to capture instantaneous velocity or pressure data The reward function is defined as rTi = CD,0 −(CD)Ti − ω(CL)Ti Sensitivity Analysis of Jet Positioning The control effectiveness is found to be highly sensitive to the location of synthetic jets, with optimal control achieved when the synthetic jets are located at the primary flow separation points At Re = 100, positioning the synthetic jets near the trailing corners is advantageous for drag reduction In contrast, at Re = 500, positioning synthetic jets near the leading corners yields a more pronounced drag reduction effect Detailed investigations of control mechanisms reveal that optimal control is achieved when the synthetic jets are located at the primary flow separation points Impact of Synthetic Jet Width The control effectiveness of drag reduction is found to be relatively insensitive to the width of synthetic jets, however, narrower jets can lead to decreased flow stability, particularly at higher Reynolds numbers At Re = 100, when the jets width decreases from D/20 to D/35, the standard deviation of the lift coefficient is reduced by 15.8% When the jet width decreases from D/20 to D/35 at Re = 500, the standard deviation of the drag coefficient is reduced by 44.5%, and the standard deviation of the lift coefficient is reduced by 33.9% Control Strategy Dynamics The overall control strategy demonstrates the ability of DRL control in achieving drag reduction while minimizing energy consumption under various flow conditions In the beginning, a large jet velocity is often used to rapidly intervene and adjust the flow state, subsequently a smaller synthetic jet velocity is employed to sustain the flow control effect and reduce energy consumption The time history curves of CD, CL, and control actions for the square cylinder at Re = 100 and Re = 500 are depicted in Fig.
Kai: 17 DRL-based AFC offers significant advantages in handling complex nonlinear systems and high-dimensional state spaces, providing powerful tools for flow control Investigating the impact of the position and width of synthetic jets on flow control performance can further optimize control strategies, enhance flow stability, and provide valuable insights for practical engineering applications The DRL agent is capable of observing not only the flow field information around a square cylinder but also more importantly, the flow field information in the most pulsating regions of the wake field Acknowledgments and Resources The authors would like to express their gratitude to Dr.
Paper summary: Kai: To pick up where we left off, this paper dives into the details of optimizing the synthetic jet actuator parameters within a deep reinforcement learning framework to control flow around a square cylinder. They are specifically investigating how the placement and width of these jets influence drag reduction and vortex suppression.
Mira: The thesis is that by using deep reinforcement learning, you can actively adjust the mass flow rates of two synthetic jets to achieve significant reductions in both lift and drag coefficients on the cylinder, while also completely suppressing vortex shedding in the wake.
Lev: So they aren't just running a simulation; they are training an agent that learns the optimal control policy based on observing the flow field around two hundred one probes placed all over the domain.
Kai: That interaction is what makes it deep reinforcement learning, right? The agent observes the current state of the CFD environment and then adjusts its actions, which are defined by dimensionless mass flow rates Q one and Q two, with those two summing to zero to keep the net mass flow at nothing.
Mira: They define their success through a reward function that measures how much they reduce drag coefficient CD,zero minus the current drag coefficient times T i plus some penalty term related to the lift coefficient CL times T i.
Lev: That’s a standard way to frame it—trying to maximize the desired output while minimizing penalties, which is exactly what you want when you're tuning a control system for physical hardware.
Kai: And they test this whole setup at two different Reynolds numbers of one hundred and five hundred to see how robust the performance is across varying flow conditions <ref:2405.12834#pg1>.
Mira: The main finding is that the placement matters a lot: near the trailing edge corners at Re equals one hundred gave a maximum drag reduction rate of fourteen point four percent, while near the leading edge corner at Re equals five hundred achieved a maximum optimal drag reduction effect of sixty-five point five percent <ref:2405.12834#pg1,a maximum optimal drag reduction effect of 65.5>.
Lev: That shift in optimal placement between those two Reynolds numbers is the core piece of information they are presenting for anyone looking to design this type of flow control system for real applications.
Kai: And they also looked into the width, finding that while it doesn't drastically change the total drag reduction percentage, narrower jets can introduce some instability at higher Reynolds numbers.
Mira: So, if you’re thinking about implementing this, you have to consider those specific physical parameters—the position and the width of those synthetic jets—because they dictate where the DRL agent will find its best performance.
Conclusion: Kai: So looking at the whole paper, "Effect of Synthetic Jets Actuator Parameters on Deep Reinforcement Learning-Based Flow Control Performance in a Square Cylinder," what this tells us is that combining DRL with fluid dynamics simulation gives us a powerful tool for active flow control.
Mira: It’s important to note that this isn't just some abstract theory; they actually built the framework and tested it against physical constraints like two different Reynolds numbers.
Lev: And what they did really brings up the question of how you translate these simulation results into something you can run on real hardware, because that’s where a lot of the complexity lies.
Kai: Exactly. The implication for someone listening who might not be in fluid dynamics is that if you have a complex nonlinear system, like unsteady flow around an object, deep reinforcement learning offers a way to find control strategies without having to manually calculate every single physical interaction from scratch.
Mira: They’re showing that the results are highly dependent on the specific flow regime, which means engineers can't just take one set of parameters and assume it works everywhere.
Lev: That dependence on the Reynolds number is a major caveat for real-world application; you need to know your operating conditions before you can trust those specific performance numbers they found.
Kai: And if we think about the authors, Wang Jia and Hang Xu, they’ve provided open source codes that let others try to use this DRL framework themselves with CFD codes like OpenFoam.
Mira: That opens the door for more researchers to build on this work, testing different jet designs or even trying different reward functions to see if they can push those performance numbers further.
Lev: It’s a good resource because it gives you a concrete starting point for integrating this type of AI control logic into your own fluid dynamics modeling environment.
More episodes
- 2611.14082-Two atoms with strong light-induced dipolar interactions
- 2611.14040-Characterization of BNLs 20 and 30 micrometer thick LGADs for the Electron Ion Collider
- 2611.13399-Lift-element-based dynamic network of vortex gust-airfoil interactions
- 2611.13476-Hopfion Superlattices in 3D Twisted Moir'e Light Fields
- 2611.13642-Investigation of the timing performance of double-sided p-type 3D silicon sensors produced by CNM
- 2611.13830-SiPM optical readout of a PICOSEC Micromegas precise timing detector
- 2611.14152-Where are the shortest-path nodes in spatial networks?
- 2611.14219-Sub-Ice: semi-automated mapping of basal channels from ice shelf surface elevation data
- 2611.16341-SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
- 2611.16090-Multi-UTE Myelin Fraction Mapping: Repeatability and Cross-Field-Strength Reproducibility at 1.5T and 3T