A Validated LBM Dataset and Pipeline for Surrogate Modeling of Turbulent 3D Obstructed Channel Flows

This paper introduces a rigorously validated, reproducible pipeline using a cumulant-based Lattice Boltzmann solver to generate a high-resolution dataset of 3D turbulent obstructed channel flows, establishing a standardized benchmark for evaluating and comparing neural operator surrogates in turbulence modeling.

The Gist

Imagine you are trying to teach a computer how to predict how water swirls around a rock in a fast-moving river. Doing this with traditional supercomputers is like trying to calculate every single water molecule's path by hand—it takes forever and costs a fortune in electricity.

This paper introduces a new "training gym" and a set of rules to teach computers a shortcut. Instead of calculating every molecule, the goal is to train a smart AI (a neural network) to guess the result almost instantly, while still getting the physics right.

Here is a breakdown of what the authors did, using simple analogies:

1. The Problem: The "Slow Cooker" vs. The "Microwave"

Traditional fluid simulations are like a slow cooker: they take a long time to get the perfect result, but the result is very accurate. The authors want to build a "microwave" (a neural network) that can give you a hot meal in seconds. But to build a good microwave, you need a massive library of perfect slow-cooked meals to learn from.

2. The Solution: A Rigorous "Training Gym"

The authors created a pipeline (a step-by-step assembly line) to generate this library of data.

• The Obstacles: They didn't just use simple shapes. They created 42 different "rocks" (objects like cylinders, spheres, and wedges) of various shapes and sizes.
• The Flow: They simulated water flowing around these rocks at different speeds (Reynolds numbers from 1,000 to 10,000). This is the "turbulent" zone where water gets chaotic and swirls wildly.
• The Resolution: To make sure the data is high-quality, they used a massive grid (1024 x 512 x 512). Think of this as using a 4K camera instead of a blurry phone camera to record the water. This ensures they can see the tiny, fast-moving swirls (eddies) that are crucial for accuracy.

3. The "Referee": Validating the Data

You can't just trust the computer; you have to check if it's telling the truth. The authors acted as strict referees by comparing their computer simulations against real-world experiments done by other scientists.

• The Checks: They checked specific "stats" of the flow:
  • The Wiggle Factor (Strouhal Number): How often the water wobbles behind the object.
  • The Drag: How hard the water pushes against the object.
  • The Swirls: How the turbulence breaks down.
• The Result: Their computer data matched the real-world experiments very closely (within about 6% error). This proves their "training gym" is legitimate and the data is trustworthy.

4. The First Test Run: The "Student Athletes"

Once they had the data, they tested a few different AI models (the "students") to see who could learn the best.

• The Contenders: They tried different types of neural networks, including a "Fourier Neural Operator" (which is good at seeing patterns in waves) and a "U-Net" (a type of network often used for image processing).
• The Winner: The U-Net model performed the best. It made the fewest mistakes and learned the fastest. The authors say this is just a "proof of concept" (a first try), but it shows the pipeline works.

5. What's Next?

The authors aren't done yet. They plan to:

• Compare Models: Systematically test which AI architecture is best at predicting the future flow, fixing errors, or turning low-quality images into high-quality ones.
• Check the Speed: They want to see if the AI is actually faster than the traditional supercomputer simulation.
• Get Feedback: They are asking the scientific community, "Is our way of testing these models fair? Are we measuring the right things?"

Summary

In short, the authors built a high-quality, verified dataset of 3D water flows around complex shapes. They proved their simulation method is accurate by comparing it to real experiments. They then used this data to train a few AI models, finding that one model (U-Net) is currently the best at predicting these flows. Their goal is to create a standard "benchmark" so that other scientists can fairly compare their own AI models for fluid dynamics in the future.

Note: The paper focuses strictly on the creation of this dataset, the validation of the simulation method, and the initial testing of AI models. It does not claim these models are ready for real-world engineering use yet, nor does it discuss medical or clinical applications.

Technical Summary

Technical Summary: A Validated LBM Dataset and Pipeline for Surrogate Modeling of Turbulent 3D Obstructed Channel Flows

Problem Statement
Computational Fluid Dynamics (CFD) simulations of 3D turbulent flows at moderate to high Reynolds numbers (Re) remain computationally prohibitive, even with modern GPU parallelization. While machine learning (ML) surrogate models offer a pathway to accelerate predictions, their rigorous evaluation is hindered by a lack of validated, high-quality datasets featuring complex geometries. Existing databases, such as the Johns Hopkins Turbulence Database, focus on canonical configurations (e.g., isotropic turbulence) rather than the 3D flows around complex objects relevant to engineering. Furthermore, fair comparison of state-of-the-art neural operators (e.g., Fourier Neural Operators, U-Nets) is difficult without consistent physical benchmarks and metrics that capture critical turbulent phenomena, such as the energy cascade and high-frequency dynamics.

Methodology
The authors present a reproducible pipeline designed to generate training data for 3D channel flows around procedurally generated geometries at Reynolds numbers ranging from 1,000 to 10,000. The methodology consists of three primary stages:

1. Geometry and Data Generation: The pipeline generates derivative objects (cylinders, rectangles, spheres, tori, wedges) within a computational domain of $2 \times 1 \times 1$ (channel height units). Simulations are conducted using the waLBerla framework, which provides highly optimized, code-generated Lattice Boltzmann Method (LBM) implementations.
2. Solver Configuration: The LBM solver utilizes a D3Q27 stencil with cumulant collision operators, offering improved numerical stability and accuracy over standard multiple relaxation time models. To enhance precision for turbulent flows, a fourth-order advection-diffusion correction scheme with limiting is employed. Notably, the approach captures the energy cascade without explicit subgrid-scale (SGS) models (e.g., Smagorinsky), as testing indicated that adding SGS models reduced physical accuracy in this regime. Fluid-solid interactions use a quadratic bounce-back scheme for second-order spatial accuracy.
3. Validation Protocol: The dataset is rigorously validated against experimental measurements (specifically Choi and Park [13] for sphere flow) and numerical benchmarks. Validation metrics include:
   • Drag Coefficients: Compared against Roos and Willmarth [15] across various grid resolutions to ensure grid convergence.
   • Strouhal Number: Validating vortex shedding dynamics.
   • Recirculation Length: Measuring the size of the recirculation bubble.
   • Turbulent Fluctuations: Assessing streamwise turbulent intensity and velocity fluctuation statistics.
   • Turbulent Structures: Visual verification of coherent structures like hairpin vortices.

The final dataset comprises flows around 42 procedurally generated objects, stored as HDF5 files at resolutions ranging from $64^3$ to $1024 \times 512 \times 512$ (simulations run at full resolution, stored at half).

Key Contributions

• Validated Pipeline: A complete, reproducible workflow for generating 3D turbulent flow data around complex geometries, bridging the gap between canonical databases and engineering-relevant scenarios.
• Rigorous Verification: The LBM solver configuration is verified against experimental data for Strouhal numbers, drag coefficients, and turbulent fluctuations, with comprehensive grid convergence studies confirming numerical accuracy at high resolutions ($1024 \times 512 \times 512$).
• Standardized Benchmarking: The work establishes a framework for the standardized comparison of neural operators (FNO, U-Net) on tasks including forecasting, super-resolution, and error correction, using physics-informed metrics.
• Preliminary Baselines: The authors provide initial performance benchmarks for Fourier Neural Operators (FNO) and U-Net architectures predicting time-averaged flow fields from geometry masks.

Results

• Solver Accuracy: The LBM solver demonstrated convergent behavior in drag coefficients, matching experimental values within $\pm 6\%$ at a resolution of 512 cells. Recirculation lengths ($l_r/D \approx 1.8$) and Strouhal numbers ($St \approx 0.18$) aligned with literature values. Instantaneous snapshots successfully reproduced fine-scale coherent structures, such as hairpin vortices.
• Surrogate Model Performance: In preliminary baseline tests predicting mean streamwise velocity fields:
  • U-Net (5 layers) achieved the lowest error rates (MSE: $1.22 \times 10^{-4}$, NRMSE: 0.0159) and the highest throughput (32.87 samples/s).
  • FNO variants (Original, Hybrid, Factorized) showed higher errors (MSE ranging from $6.01 \times 10^{-4}$ to $11.13 \times 10^{-4}$) and lower throughput compared to the U-Net.
  • The U-Net is identified as the most promising candidate for further research based on these initial results.

Significance and Future Work
The paper positions this work as a foundational step toward advancing physics-based neural networks and hybrid differentiable numerical solvers. By providing a validated dataset and pipeline, the authors aim to enable fair, systematic comparisons of neural operators in turbulent flow regimes.

The authors explicitly state that the current baseline results serve only as proof-of-concept demonstrations, lacking comprehensive hyperparameter tuning and variance estimation. Future work outlined in the paper includes:

• Systematic evaluation of U-Net and FNO variants on forecasting, super-resolution, and error correction tasks.
• Implementation of expanded loss functions combining adversarial and spectral components to address spectral bias (the tendency to oversmooth fine-scale structures).
• Exploration of attention mechanisms in Time-Conditioned U-Nets and FNOs.
• A rigorous comparison of computational costs between traditional solvers and neural surrogates to assess practical applicability.

The authors invite community feedback on their validation approach and evaluation priorities, emphasizing that the ultimate goal is to support the development of efficient, physically accurate surrogate models for 3D turbulent flows.