Vortex shedding suppression in elliptical cylinder via reinforcement learning

arXiv:2410.00421 · physics.flu-dyn · Submitted 2024-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Vortex shedding suppression in elliptical cylinder via reinforcement learning".

Mira: Here's the gist of the paper. Flow control of bluff bodies plays a critical role in engineering applications In this study, deep reinforcement learning (DRL) is employed to develop flow control strategies for the flow past an elliptical cylinder confined between two walls DRL training results demonstrate that for an elliptical cylinder with larger Ar, the control strategy effectively reduces drag, minimizes lift fluctuations, and completely suppresses vortex shedding, all while maintaining low external energy consumption Through detailed physical analysis, the coupling effect between the blockage ratio (beta) and Ar is identified as a limiting factor for vortex shedding suppression and wake stabilization Pruning redundant reasoning tokens The primary objective of this work is to investigate whether it is possible to manipulate the synthetic jet based on the DRL algorithm to achieve efficient, stable and energy-saving flow control effects in the case of blunt-body flow around an elliptical cylinder with an extreme slenderness ratio The research focuses on real-time adaptive control techniques for high-dimensional nonlinear flow systems, with an emphasis on improving energy efficiency while maintaining control performance How the training signal is built The DRL framework integrates DRL with CFD to address the problem of AFC, conceptualizing fluid dynamics simulation as an interactive environment The agent interacts with the environment through state observations si, actions ai, and reward signals ri The state space S involves extracting measurements at probe points from CFD simulations to capture pronounced instability The action space A is defined by the dimensionless mass flow rates Q1 and Q2, constrained by Q1 + Q2 = 0 to ensure zero net mass flow The reward function rt is designed to minimize the drag coefficient while introducing a penalization term based on the lift coefficient to discourage unsteady or asymmetric control behaviors Where the gains come from The effectiveness of the DRL-based flow control strategy was highly sensitive to both Ar and beta of the elliptical cylinder With a relatively high blockage ratio of beta = 0.24, effective suppression of vortex shedding was achieved for configurations with Ar = 1.0, 0.75, and 0.5, with the control input requiring only about 1% of the inlet flow energy In contrast, for more slender bodies with Ar = 0.25 and 0.1, the DRL control algorithm managed to achieve only a modest level of drag reduction When the blockage ratio beta was reduced to beta = 0.12, the DRL framework showed improved adaptability and robustness across all tested Ar values Analysis of incomplete vortex suppression Comparative instantaneous vorticity contours for elliptical cylinders with varying Ar show that under controlled flow conditions, the wake exhibits a stable and symmetric distribution of positive and negative vorticity, indicating that vortex shedding has been successfully mitigated However, as the Ar decreases to 0.5, the suppression of vortex shedding becomes increasingly challenging For even more slender elliptical cylinders with Ar = 0.25 and Ar = 0.1, pronounced periodic vortex shedding remains clearly observable before and after control implementation Energy consumption trade-off The results highlight a trade-off between control effectiveness and energy expenditure, which varies significantly across different Ar configurations For the cylinder with Ar = 1, the application of the DRL-based flow control strategy leads to a 7.9% reduction in drag, while maintaining an average lift coefficient of-0.041 and external energy consumption amounting to only 0.1% of the incoming flow However, as the Ar decreases further, although the drag reduction rate improves, the associated external energy consumption increases significantly For Ar = 0.5, 0.25, and 0.1, the energy consumption ratio reaches up to 14.5% This study demonstrated that DRL had strong potential in adapting to varied flow conditions, offering a data-driven solution for real-time flow control over a wide range of bluff body geometries at different aspect and blockage ratios Future research could integrate real-time sensor feedback and extend the control framework to more complex and unsteady flow environments The learned policies led to substantial drag reduction, ranging from 6.1% to 32.3%, while maintaining near-zero lift fluctuations across all configurations The study confirms that as the aspect ratio decreased, achieving stable and energy-efficient control became increasingly challenging The DRL framework showed improved adaptability and robustness across all tested Ar values when the blockage ratio beta was reduced to beta = 0.12 The study highlights a trade-off between geometry and control efficiency: as the aspect ratio decreased, achieving stable and energy-efficient control became increasingly challenging The DRL framework showed improved adaptability and robustness across all tested Ar values when the blockage ratio beta was reduced to beta = 0.

Title and authors: Kai: Moving on to how they structured this research, we have a look at the title and authors of this work, "Vortex shedding suppression in elliptical cylinder via reinforcement learning." It sets the stage for what they are trying to achieve.

Mira: The paper is clearly focused on using deep reinforcement learning to actively suppress vortex shedding around an elliptical cylinder. This isn't just a static fluid dynamics problem; it’s about creating a dynamic control system.

Lev: From an engineering viewpoint, that implies they’re developing a policy that can make continuous adjustments to the flow control inputs based on what the flow is doing right now. It’s not about finding one perfect solution for one cylinder shape; it's about learning how to handle variations.

Kai: Right, and the authors are looking at this feasibility question: can we actually achieve multi-objective flow control—reducing drag while keeping energy input low—for an elliptical cylinder that changes its aspect ratio? That’s a tough goal.

Mira: They found that the DRL training results demonstrate that for a larger aspect ratio, say one point zero, the control strategy effectively reduces drag and minimizes lift fluctuations while keeping external energy consumption low. But when you decrease the aspect ratio, it gets harder.

Lev: That’s where they identified the coupling effect between the blockage ratio and the aspect ratio as a limiting factor for both vortex shedding suppression and wake stabilization. It’s not just one variable that matters here; it's how they interact with each other.

Kai: So, to summarize this part, the paper is about developing DRL strategies for flow control of bluff bodies, specifically focusing on elliptical cylinders at varying aspect ratios while trying to keep the energy input reasonable.

Mira: It’s a validation that data-driven modeling can find these multi-objective solutions where traditional methods might struggle with the complexity of those nonlinear flow systems.

Lev: It suggests that for real hardware, we need models that are robust enough to handle those geometric variations because if the policy only works well for one specific shape, it’s not practical.

The paper's summary: Kai: Now let’s look at what they actually found in the summary of "Vortex shedding suppression in elliptical cylinder via reinforcement learning." They confirm that DRL is being used to develop flow control strategies for the flow past an elliptical cylinder confined between two walls.

Mira: The core idea they are exploring is how this deep reinforcement learning agent interacts with the CFD simulations, treating fluid dynamics as an interactive environment. The agent learns through state observations, actions, and reward signals.

Lev: So when I think about that from a hardware perspective, those state observations are essentially the measurements taken from the CFD—the data that tells the agent what kind of flow instability it's seeing.

Kai: And then the action space is defined by two dimensionless mass flow rates, Q one and Q two, with the key constraint being that their sum has to be zero to ensure no net mass flow is injected or removed.

Mira: The reward function they designed is what drives the agent's learning: it’s engineered to minimize the drag coefficient while adding a penalization term based on the lift coefficient so it steers away from unsteady or asymmetric control behaviors.

Lev: That makes sense; they are explicitly trying to use reinforcement learning as an active flow control method where the agent learns optimal actuation strategies directly from the simulation environment.

Kai: And what they found is that this integrated DRL framework significantly bolsters the efficiency and effectiveness of fluid dynamics problem-solving when applied to active flow control.

Mira: The summary highlights that machine learning aids in identifying low-dimensional manifolds and flow states, which helps in a deeper understanding of fluid behavior than traditional empirical laws alone.

Lev: It’s about using the AI to find those shortcuts through the high dimensionality of the flow problem, which is a huge leap for complex simulations.

The paper's improvements: Kai: Let’s talk about what improvements this paper suggests they made to their approach in "Vortex shedding suppression in elliptical cylinder via reinforcement learning." They are discussing advancements in how DRL is applied to Active Flow Control.

Mira: They mention advancements from other research, like applying Soft Actor-Critic algorithms for square cylinders, or a dynamic RL scheme designed to reduce drag under partial measurements, or even feature-based DRL algorithms that handle sparse sensor data.

Lev: That shows they aren't just reinventing the wheel; they’re building on existing machine learning techniques from other fluid dynamics problems, which makes sense because those foundational models are already proven in some contexts.

Kai: They also highlight advancements made through optimizing probe placement and quantity, as well as sensitivity analyses of jet positioning. These refinements show that the way you measure the flow is just as important as the control itself.

Mira: And they also point to extensions into alternative actuation methods, like transitioning from synthetic jets to other ways of controlling the flow or extending DRL-based control to slender structures, which shows versatility in their technique.

Lev: From a hardware standpoint, that’s telling me that this isn't just a one-off setup; it’s adaptable enough to be considered a versatile tool for different types of flow problems.

Conclusion: Kai: So, let’s wrap up with the conclusion of "Vortex shedding suppression in elliptical cylinder via reinforcement learning." They summarize the paper's main implications and look ahead.

Mira: The main implication is that DRL has strong potential in adapting to varied flow conditions, offering a data-driven solution for real-time flow control over a wide range of bluff body geometries at different aspect ratios.

Lev: They found that the learned policies led to substantial drag reduction, ranging from six point one percent to thirty two point three percent while maintaining near-zero lift fluctuations across all configurations.

Kai: And they also confirmed that as the aspect ratio decreased, achieving stable and energy-efficient control became increasingly challenging. That’s a key constraint you need to keep in mind for any design choices.

Mira: The paper confirms that the DRL framework showed improved adaptability and robustness across all tested aspect ratios when they reduced the blockage ratio to zero point one two.

Lev: I just think that this whole study, "Vortex shedding suppression in elliptical cylinder via reinforcement learning," highlights a clear trade-off between geometry and control efficiency: as the aspect ratio decreased, stable and energy-efficient control got harder to achieve.

Kai: So it’s about using DRL to find solutions for flow control over a wide range of bluff body geometries, but remember that the performance depends heavily on those geometric parameters.

Mira: And they're pointing toward future work involving real-time sensor feedback and extending the control framework to more complex and unsteady flow environments. That’s where things are headed next.

Wang Jia, Hang Xu

State Key Laboratory of Ocean Engineering, School of Ocean and Civil Engineering, Shanghai Jiao Tong University

physics.flu-dyn

Submitted: 2024-10-01

Updated: 2025-07-21

Journal ref: International Journal of Mechanical Sciences 302, 110517 (2025)

DOI: 10.1016/j.ijmecsci.2025.110517

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: Flow control of bluff bodies plays a critical role in engineering applications In this study, deep reinforcement learning (DRL) is employed to develop flow control strategies for the flow past an…

Key concepts

Deep Reinforcement Learning (DRL)
DRL is an AI method where an agent learns to make decisions by interacting with an environment. In this study, the agent adjusts synthetic jet mass flow rates based on real-time CFD simulation data to achieve desired flow control goals like reducing drag and stopping vortex shedding.
Synthetic Jet
A synthetic jet is a device that creates a small, pulsed jet of fluid from a surface. The DRL framework controls the mass flow rates of these jets to actively manipulate the surrounding fluid flow around the elliptical cylinder for better control.
Blockage Ratio ($eta$)
The blockage ratio measures how much space is taken up by the cylinder relative to its surroundings. This parameter, along with the aspect ratio ($Ar$), was found to be a limiting factor in effectively suppressing vortex shedding and stabilizing the wake.
Aspect Ratio ($Ar$)
The aspect ratio describes the shape of the elliptical cylinder, relating its length to its width. The study showed that control effectiveness varied significantly depending on $Ar$, with better performance observed for larger $Ar$ values.

Terminology

Summary

Flow control of bluff bodies plays a critical role in engineering applications In this study, deep reinforcement learning (DRL) is employed to develop flow control strategies for the flow past an elliptical cylinder confined between two walls DRL training results demonstrate that for an elliptical cylinder with larger Ar, the control strategy effectively reduces drag, minimizes lift fluctuations, and completely suppresses vortex shedding, all while maintaining low external energy consumption Through detailed physical analysis, the coupling effect between the blockage ratio (beta) and Ar is identified as a limiting factor for vortex shedding suppression and wake stabilization

Pruning redundant reasoning tokens

The primary objective of this work is to investigate whether it is possible to manipulate the synthetic jet based on the DRL algorithm to achieve efficient, stable and energy-saving flow control effects in the case of blunt-body flow around an elliptical cylinder with an extreme slenderness ratio The research focuses on real-time adaptive control techniques for high-dimensional nonlinear flow systems, with an emphasis on improving energy efficiency while maintaining control performance

How the training signal is built

The DRL framework integrates DRL with CFD to address the problem of AFC, conceptualizing fluid dynamics simulation as an interactive environment The agent interacts with the environment through state observations si, actions ai, and reward signals ri The state space S involves extracting measurements at probe points from CFD simulations to capture pronounced instability The action space A is defined by the dimensionless mass flow rates Q1 and Q2, constrained by Q1 + Q2 = 0 to ensure zero net mass flow The reward function rt is designed to minimize the drag coefficient while introducing a penalization term based on the lift coefficient to discourage unsteady or asymmetric control behaviors

Where the gains come from

The effectiveness of the DRL-based flow control strategy was highly sensitive to both Ar and beta of the elliptical cylinder With a relatively high blockage ratio of beta = 0.24, effective suppression of vortex shedding was achieved for configurations with Ar = 1. p. 00, 0.75, and 0.5, with the control input requiring only about 1% of the inlet flow energy In contrast, for more slender bodies with Ar = 0. p. 025 and 0.1, the DRL control algorithm managed to achieve only a modest level of drag reduction When the blockage ratio beta was reduced to beta = 0. p. 212, the DRL framework showed improved adaptability and robustness across all tested Ar values

Analysis of incomplete vortex suppression

Comparative instantaneous vorticity contours for elliptical cylinders with varying Ar show that under controlled flow conditions, the wake exhibits a stable and symmetric distribution of positive and negative vorticity, indicating that vortex shedding has been successfully mitigated However, as the Ar decreases to 0.5, the suppression of vortex shedding becomes increasingly challenging For even more slender elliptical cylinders with Ar = 0. p. 825 and Ar = 0.1, pronounced periodic vortex shedding remains clearly observable before and after control implementation

Energy consumption trade-off

The results highlight a trade-off between control effectiveness and energy expenditure, which varies significantly across different Ar configurations For the cylinder with Ar = 1, the application of the DRL-based flow control strategy leads to a 7.9% reduction in drag, while maintaining an average lift coefficient of -0.041 and external energy consumption amounting to only 0. p. 01% of the incoming flow However, as the Ar decreases further, although the drag reduction rate improves, the associated external energy consumption increases significantly For Ar = 0. p. 05, 0.25, and 0.1, the energy consumption ratio reaches up to 14. p. 05%

Conclusion

This study demonstrated that DRL had strong potential in adapting to varied flow conditions, offering a data-driven solution for real-time flow control over a wide range of bluff body geometries at different aspect and blockage ratios Future research could integrate real-time sensor feedback and extend the control framework to more complex and unsteady flow environments The learned policies led to substantial drag reduction, ranging from 6.1% to 32.3%, while maintaining near-zero lift fluctuations across all configurations The study confirms that as the aspect ratio decreased, achieving stable and energy-efficient control became increasingly challenging The DRL framework showed improved adaptability and robustness across all tested Ar values when the blockage ratio beta was reduced to beta = 0.12 The study highlights a trade-off between geometry and control efficiency: as the aspect ratio decreased, achieving stable and energy-efficient control became increasingly challenging The DRL framework showed improved adaptability and robustness across all tested Ar values when the blockage ratio beta was reduced to beta = 0. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Real-time adaptive flow control for complex bluff body flows can be achieved by developing a deep reinforcement learning framework that adjusts synthetic jet mass flow rates based on instantaneous CFD simulation data, as described in Section 2.3. This allows the system to continuously adapt to the changing flow characteristics in real-time and maintain a balanced trade-off between real-time adaptation and computational performance.

  2. The improved AI system can precisely predict drag reduction rates and lift coefficient suppression across a wide range of aspect ratios (Ar) and blockage ratios (β). This capability is demonstrated by the results showing that the DRL-based flow control strategy exhibits outstanding control performance with drag reductions ranging from 6.1% to 32.3%.

  3. The system can dynamically optimize energy efficiency alongside aerodynamic performance by learning the optimal actuation strategies for different geometries. For instance, it can determine that as the aspect ratio decreases, the associated external energy consumption increases significantly, allowing designers to select a control strategy that minimizes energy input while achieving target flow stability.

  4. The improved AI system will be robust against geometric variations in bluff bodies by learning policies that are sensitive to both Ar and β. This is evidenced by the finding that the effectiveness of the DRL-based flow control strategy was highly sensitive to both the Ar and the β of the elliptical cylinder.

Sources

Related papers