Vortex-Induced Drag Forecast for Cylinder in Non-uniform Inflow

This study develops a physics-based data-driven strategy using a modified fully connected neural network that integrates upstream velocity measurements with optimized pressure sensor data to accurately predict vortex-induced drag on a circular cylinder under non-uniform inflow conditions at a Reynolds number of 4000.

The Gist

Imagine you are trying to predict how hard the wind will push against a flagpole on a stormy day. If the wind were blowing perfectly straight and steady, it would be easy to guess the force. But in the real world, the wind is messy. It swirls, gusts, and changes direction unpredictably. This is what engineers call "non-uniform inflow," and it makes predicting the force on structures (like bridges, smokestacks, or underwater cables) incredibly difficult.

This paper is about a new, smart way to predict that force using a combination of physics and artificial intelligence (AI). Here is the breakdown in simple terms:

The Problem: The "Messy Wind" Puzzle

For over a century, scientists have studied how wind flows around cylinders (like poles). Usually, they try to predict the force by looking at the pressure on the surface of the pole.

• The Old Way: Think of it like trying to guess the weather inside a house just by listening to the wind outside the front door. If the wind is steady, this works. But if the wind is chaotic and swirling (turbulent), the sound at the door doesn't tell you enough about the chaos happening in the living room.
• The Failure: In this study, when the researchers tried to use only pressure sensors on the pole to predict the force in "messy" wind conditions, their AI model was basically guessing. It had zero accuracy. The complex swirls of the wind broke the connection between the pressure on the pole and the total force pushing it.

The Solution: The "Upstream Calibrator"

The researchers realized that to understand the mess inside the wind, they needed to know what the wind looked like before it hit the pole.

• The Analogy: Imagine you are a chef trying to bake a cake. If you only taste the batter at the end, you might not know if the oven temperature was wrong. But if you check the oven temperature before you put the cake in, you can predict how the cake will turn out.
• The Fix: They added upstream velocity sensors. These are like thermometers placed in the wind before it hits the pole. They tell the AI, "Hey, the wind is about to get really gusty," so the AI can adjust its prediction of the force on the pole.

The AI "Brain" and the Optimization Game

They built a neural network (a type of AI brain) to do the predicting. But they had a problem: Where exactly should they put the sensors?

• The Sensor Hunt: There were 32 possible spots to put pressure sensors around the pole. Trying every combination of sensors would take forever (like trying every possible combination of keys on a giant piano to find the right song).
• The Strategy: They used a smart, step-by-step optimization process. They started with one sensor, saw how well it worked, added another, and kept going.
• The Discovery: They found that they didn't need 32 sensors. Just a few, placed in specific spots, worked wonders.
  • The Magic Spots: The best sensors were placed right where the wind starts to peel away from the pole (the "separation points"). It's like realizing that to predict a traffic jam, you don't need to watch every car; you just need to watch the two cars where the lane first narrows.
  • The "Exponential" Rule: They found a cool pattern: Adding the first few sensors gave a huge boost in accuracy. Adding more sensors after that gave smaller and smaller improvements. It's like filling a bucket: the first cup of water fills it up 50%, the second cup fills it another 30%, but the tenth cup only adds a tiny splash.

The Results: A Crystal Ball for Engineers

The final model was a huge success:

1. Accuracy: It went from being useless (0% accuracy) to being very good (75% accuracy) at predicting how much the wind would push the pole in the next second.
2. Speed: The model is fast enough to run in real-time.
3. Robustness: Even when the wind was extremely chaotic, the model could still tell engineers, "Hey, a big gust is coming," allowing them to reinforce structures before they get damaged.

Why This Matters

This isn't just about poles in the wind. This is a blueprint for how to use AI in engineering when the environment is messy and unpredictable.

• Real-world impact: This could help design safer bridges, more efficient wind turbines, and sturdier offshore oil rigs.
• The Big Lesson: You don't need to measure everything to predict the future. You just need to measure the right things (the upstream wind) and place your sensors in the right spots (where the flow separates).

In short, the authors taught a computer to look at the wind before it hits the object, combined with a few cleverly placed sensors, to predict the chaos that follows. It's a smarter, more efficient way to keep our engineering structures safe in a turbulent world.

Technical Summary

1. Problem Statement

The paper addresses the challenge of predicting vortex-induced drag on a circular cylinder under non-uniform inflow conditions, a scenario common in real-world engineering (e.g., ocean and wind engineering) at moderate Reynolds numbers ($Re \approx 4000$).

• Limitations of Existing Methods: Traditional data-driven models relying solely on surface pressure signals (often processed via wavelet transforms) fail in non-uniform inflows. The complex vortex dynamics, three-dimensional wake instabilities, and spatially heterogeneous pressure distributions caused by turbulence weaken the direct correlation between surface pressure and drag force.
• The Gap: Pure pressure-signal-based models exhibit near-zero predictive capability ($R^2 \sim 0$) in these conditions because they cannot resolve the multi-scale frequency peaks introduced by the non-uniform inflow.

2. Methodology

The authors propose a physics-based data-driven strategy that integrates upstream velocity measurements with surface pressure signals to calibrate the inflow conditions.

• Numerical Setup:

  • Simulation: Direct Numerical Simulations (DNS) of flow over a circular cylinder at $Re = 4000$ using the spectral element code Nek5000.
  • Inflow: A non-uniform inflow is generated using periodic boundary conditions in the streamwise, spanwise, and cylinder-parallel directions.
  • Data: Time-resolved velocity and pressure data were collected from 32 points along the cylinder surface and 3 upstream points located at $(-2,0,0)$, $(-1,0,0)$, and $(-0.7,0,0)$.
  • Signal Processing: Pressure signals were processed using Discrete Wavelet Transform (DWT) to extract characteristic frequency coefficients ($\gamma$) and their temporal derivatives ($d\gamma/dt$).

• Machine Learning Architecture:

  • Model: A modified Fully Connected Neural Network (FCNN) with the architecture: Input $\to$ FC24 $\to$ FC28 $\to$ FC24 $\to$ FC16 $\to$ FC8 $\to$ Output.
  • Input Strategy: Instead of using raw pressure, the model uses DWT-processed coefficients. Crucially, the input array is augmented with upstream velocity components ($u, v, w$) and their time derivatives to serve as an "inflow calibration."
  • Optimization: An iterative optimization process was conducted to determine:
    1. The optimal placement of pressure sensors on the cylinder surface.
    2. The optimal combination of upstream velocity components.
  • Training: The model was trained to predict the drag coefficient ($C_d$) at a future time window ($\Delta t_p = 1 D/U$) using a Mean Absolute Error (MAE) loss function and AdamW optimizer.

3. Key Contributions

• Inflow Calibration Strategy: The study demonstrates that incorporating upstream velocity measurements acts as a physical calibration for non-uniform inflows, resolving the spectral mismatch that causes pure pressure models to fail.
• Sensor Placement Optimization: Through iterative selection, the authors identified specific optimal locations for pressure sensors that maximize predictive accuracy with minimal sensors.
• Exponential Scaling Law: A novel observation was made where model performance ($R^2$) scales exponentially with the number of optimized pressure signals added to the input, asymptotically approaching a maximum performance limit ($R^2_b$).
• Physical Interpretation: The optimized sensor locations correspond to the leading edges of flow separation points, validating the physical mechanism that flow separation dynamics govern vortex-induced drag generation.

4. Key Results

• Performance Improvement:
  • Pure pressure-based models achieved $R^2 \sim 0$.
  • The optimized hybrid model (upstream velocity + optimized pressure) achieved an $R^2$ score of 0.75.
  • The model successfully forecasts high-amplitude drag fluctuations ($C_d$ ranging from 0.2 to 1.2).
• Optimal Configurations:
  • Pressure Sensors: The most effective sensors were located at $pt = 18, 30, 8, 22$ (using velocities at $(-2,0,0)$). These locations align with flow separation leading edges and trailing edges.
  • Velocity Inputs: The streamwise velocity ($u$) provided the most information, followed by the transverse velocity ($v$). The spanwise component ($w$) was negligible. The optimal input used velocities at $(-2,0,0)$ and $(-1,0,0)$.
• Scalability:
  • The model maintains qualitative consistency even with extended prediction windows ($\Delta t_p = 5 D/U$), though accuracy decreases monotonically as the window increases.
  • A small number of sensors (e.g., 2 pressure signals + specific velocity components) can achieve over 90% of the maximum possible performance ($R^2_b$).
• Validation: The predicted $C_d$ time series and probability density functions (PDFs) closely matched the DNS benchmarks, capturing both the mean trends and the chaotic fluctuations.

5. Significance

• Engineering Application: This work provides a scalable, computationally efficient strategy for real-time drag forecasting in turbulent, non-uniform flow environments, which is critical for preventing structural damage (e.g., fatigue, lock-in vibrations) in offshore and wind engineering.
• Physics-Informed ML: It bridges the gap between pure data-driven approaches and physical understanding. By interpreting the optimal sensor locations through fluid dynamics (flow separation), the study validates that machine learning can uncover governing physical mechanisms when guided by appropriate inputs.
• Future Potential: The proposed "inflow calibration" strategy offers a pathway for experimental validation, active flow control schemes, and the development of reduced-order models for complex fluid-structure interaction systems.

Conclusion:
The paper successfully overcomes the limitations of pressure-only models in non-uniform flows by integrating upstream velocity data. It establishes a robust, physics-informed machine learning framework that achieves high predictive accuracy ($R^2 = 0.75$) with a minimal sensor footprint, offering a practical solution for monitoring and controlling vortex-induced vibrations in engineering applications.