WAKESET: A Large-Scale, High-Reynolds Number Flow Dataset for Machine Learning of Turbulent Wake Dynamics

This paper introduces WAKESET, a novel large-scale dataset comprising over 4,000 high-fidelity simulations of high-Reynolds number turbulent flows during underwater vehicle recovery, designed to overcome data scarcity and advance machine learning applications in computational fluid dynamics.

The Gist

Imagine you are trying to teach a robot how to swim through a stormy ocean, but the only way to learn is by actually swimming in the storm. The problem? The storm is dangerous, the ocean is huge, and simulating it on a computer takes so much time and energy that you could only ever practice a few times before your computer burns out.

This is the exact problem engineers face with Computational Fluid Dynamics (CFD). They need to understand how water (or air) moves around complex objects like submarines or drones, but running the super-computer simulations required to get accurate data is incredibly expensive and slow. It's like trying to learn to drive a Formula 1 car by only being allowed to drive it once a year.

Enter "WAKESET."

Think of WAKESET as the "ImageNet for underwater physics." Just as the famous ImageNet dataset helped computers learn to recognize cats and dogs by showing them millions of pictures, WAKESET is a massive library of underwater flow simulations designed to teach Artificial Intelligence (AI) how to understand water movement.

Here is the story of how they built it, explained simply:

1. The Big Challenge: The "Data Desert"

In fields like computer vision (recognizing images), AI has massive datasets to learn from. In fluid dynamics, it's a desert. Most existing datasets are like looking at a flat, 2D drawing of a river. They are too small, too simple, or only show slow-moving water. Real-world engineering, however, involves turbulent, high-speed, 3D chaos.

To train a smart AI that can actually help engineers design better submarines, they needed a dataset that was:

• Huge: Thousands of examples, not just a handful.
• Fast: Simulating water moving at very high speeds (high Reynolds numbers).
• 3D: Showing the full volume of water, not just a slice.

2. The Case Study: The Underwater "Docking"

To make this dataset useful, the researchers picked a very specific, tricky real-world problem: A small underwater drone (AUV) trying to dock inside a giant underwater mothership (XLUUV).

Imagine a tiny submarine trying to sneak into the cargo bay of a massive submarine while both are moving through the ocean.

• The Chaos: The big ship creates a massive wake (a turbulent trail of swirling water) behind it.
• The Danger: The small drone has to navigate through this swirling mess, dealing with spinning water, pressure changes, and the big ship's propeller wash.
• The Goal: The AI needs to learn exactly how the water behaves so it can predict where the drone will be pushed and how to steer it safely.

3. Building the Library: From One to Thousands

The researchers didn't just run one simulation. They built a "recipe book" for the water:

• The Generalized Ship: Instead of modeling one specific ship, they created a "generic" giant underwater ship that represents the average design of these massive vessels. This ensures the AI learns the principles of water flow, not just the quirks of one specific boat.
• The Variable Menu: They ran simulations with the ship moving at different speeds (from a slow crawl to a sprint) and turning at different angles.
• The Magic Trick (Data Augmentation): Running a simulation is expensive. To get more data without spending more money, they used a clever trick. If they simulated the ship turning left, they mathematically "flipped" the data to create a right turn. If they simulated a straight path, they mirrored it.
  • Analogy: It's like taking one photo of a person and using a mirror to create a photo of them facing the other way. You didn't take a new photo, but you now have two different angles to study.
  • Result: They started with 1,091 simulations and "augmented" them into 4,364 unique training examples.

4. What's Inside the Box?

The WAKESET dataset is a massive 480GB collection of data. It doesn't just say "the water is moving." It provides a 3D grid of data points showing:

• How fast the water is moving in every direction.
• The pressure pushing on the ship.
• How "swirly" (turbulent) the water is.
• How the wake changes as the ship turns.

5. Why Does This Matter?

Before WAKESET, if an engineer wanted to design a new underwater vehicle, they had to wait weeks for a supercomputer to simulate the water flow for every tiny change they made.

With WAKESET, they can train an AI model. Once trained, this AI can act as a "Super-Predictor."

• Old Way: Wait 3 days for a computer to calculate the water flow.
• New Way: The AI looks at the design and predicts the water flow in milliseconds.

This allows engineers to test thousands of designs instantly, optimize them for safety and speed, and even create real-time control systems for autonomous drones that can react to underwater currents instantly.

The Bottom Line

WAKESET is the fuel for the next generation of underwater AI. By providing a massive, high-quality library of how water behaves in complex, real-world scenarios, it allows machine learning models to finally "learn" the physics of the ocean. It bridges the gap between slow, expensive computer simulations and the fast, smart AI needed to explore the deep sea.

Technical Summary

1. Problem Statement

The integration of Machine Learning (ML) into Computational Fluid Dynamics (CFD) is hindered by a critical scarcity of large-scale, high-fidelity, and diverse datasets. Existing resources suffer from three main limitations:

• Scale and Diversity: Most datasets contain only hundreds of instances, often focusing on canonical flows, low Reynolds numbers, or 2D simplifications. This is insufficient for training deep learning architectures that require massive data to prevent overfitting and ensure generalization.
• Physical Fidelity: Many datasets lack the complexity of real-world engineering problems, specifically high-Reynolds number turbulent flows (e.g., $Re > 10^7$) and fully 3D flow fields.
• Computational Cost: Generating high-fidelity data (DNS or high-quality RANS/LES) for complex geometries is computationally prohibitive, limiting the ability to create the "ImageNet" equivalent for fluid mechanics.

Consequently, ML models struggle to generalize to practical engineering scenarios involving complex 3D turbulence, such as underwater vehicle maneuvering and recovery.

2. Methodology

The authors developed WAKESET through a rigorous two-phase approach, centered on the hydrodynamic challenge of an Autonomous Underwater Vehicle (AUV) being recovered inside the payload bay of an Extra-Large Uncrewed Underwater Vehicle (XLUUV).

Phase 1: Foundational Hydrodynamic Analysis

• Case Study: A specific scenario involving an AUV docking with an XLUUV was analyzed to identify critical flow phenomena (shear layers, propeller wakes, recirculation zones).
• Validation: High-fidelity Reynolds-Averaged Navier-Stokes (RANS) simulations were performed using ANSYS Fluent with a realizable $k-\epsilon$ turbulence model.
• Outcome: This phase validated the CFD methodology and identified the specific flow regions (e.g., payload bay entrance, wake) requiring high mesh resolution.

Phase 2: Generalization and Dataset Creation

To transform the specific analysis into a versatile ML training resource, the authors generalized the model and expanded the parameter space:

• Geometry Generalization: A representative XLUUV geometry (22m length, 2.2m x 2.7m beam) with a standardized internal payload bay was created. This avoids overfitting to a single specific vehicle design while capturing essential hydrodynamic interactions.
• Parametric Expansion:
  • Speed: Forward speeds varied from 0.10 m/s to 5.00 m/s.
  • Maneuvering: Turning angles (yaw) varied from 0° to 60°.
  • Reynolds Number: The dataset covers Reynolds numbers up to $1.09 \times 10^8$, among the highest in public CFD datasets.
• Simulation Scale: 1,091 unique RANS simulations were generated on the GADI supercomputer.
• Data Augmentation: To increase dataset size and diversity without new simulations, the authors applied:
  • Rotation: Rotating flow fields to simulate opposite turning directions.
  • Flipping: Mirroring symmetric 0° cases across the vertical mid-plane.
  • Result: The dataset was expanded from 1,091 to 4,364 instances.

Data Structure

The dataset includes:

• 3D Volumes: Interpolated onto a $128^3$ Cartesian grid (optimized for volumetric ML like 3D-GANs).
• 2D Planes: Unstructured mesh data on vertical and horizontal slices (retaining original CFD density for boundary layer analysis).
• Variables: Velocity components ($u, v, w$), velocity magnitude, pressure (static, dynamic, total), vorticity, and turbulence intensity.

3. Key Contributions

1. WAKESET Dataset: A publicly available, large-scale CFD dataset containing 4,364 high-fidelity 3D flow field instances.
2. Record-Breaking Scale: It features the highest upper Reynolds number ($1.09 \times 10^8$) and one of the largest instance counts (4,364) among publicly available CFD datasets, surpassing existing resources like JHTDB or DrivAerML in terms of operational envelope diversity.
3. Practical Engineering Focus: Unlike canonical flow datasets, WAKESET addresses a complex, real-world problem (AUV recovery within an XLUUV), capturing asymmetric wakes, payload bay recirculation, and maneuver-induced vortices.
4. Benchmarking Framework: The authors established a baseline for ML performance using state-of-the-art Generative Adversarial Networks (GANs) on both 2D slice prediction and 3D volumetric generation tasks.

4. Results and Performance Benchmarks

The authors benchmarked the dataset using several GAN architectures (cDCGAN, SAGAN, PatchGAN, WGAN-GP, 2D3DGAN) to predict flow fields based on input kinematic parameters (speed and turning angle).

• 2D Prediction (512x512 slices):
  • cDCGAN and PatchGAN achieved high fidelity with low relative error in area-averaged kinetic energy ($\epsilon_{Ek} \approx 1.5\% - 1.8\%$) and high Structural Similarity (SSIM > 0.99).
  • SAGAN underperformed, suggesting self-attention mechanisms may be less efficient for constrained 2D tasks compared to convolutional approaches.
• 3D Prediction (128x128x128 volumes):
  • SAGAN emerged as the most physically accurate model for 3D tasks, achieving the lowest kinetic energy error ($\epsilon_{Ek} = 8.1\%$) and best Fréchet Inception Distance (FID), demonstrating the value of self-attention in capturing volumetric flow structures.
  • 2D3DGAN offered the best computational efficiency (fastest inference time) with strong structural similarity.
• General Findings: The benchmarks confirmed that while simple architectures work well for 2D slices, capturing complex 3D turbulent wakes at high Reynolds numbers requires specialized architectures (like SAGAN or 2D3DGAN) capable of handling high dimensionality and asymmetry.

5. Significance

• Bridging the Gap: WAKESET addresses the "data scarcity" bottleneck in fluid mechanics ML, providing a resource comparable to ImageNet in computer vision.
• Enabling Real-World Applications: By focusing on high-Reynolds number, 3D turbulent flows, the dataset enables the development of ML models that can generalize to full-scale engineering problems, such as autonomous underwater navigation and control.
• Accelerating CFD: The dataset supports the creation of "surrogate models" that can predict flow fields orders of magnitude faster than traditional CFD, facilitating real-time control, design optimization, and uncertainty quantification.
• Community Resource: The dataset is open-source (480GB), includes Python/PyTorch loaders, and provides detailed documentation, fostering reproducibility and collaboration across fluid dynamics and AI research communities.

In conclusion, WAKESET represents a significant leap forward in data-driven fluid dynamics, providing the necessary scale, fidelity, and complexity to train the next generation of ML models for turbulent flow prediction and control.