TLDR: ChannelFlow-Tools is a new framework that standardizes the entire process of creating 3D obstructed channel flow datasets for machine learning. It covers everything from generating complex obstacle geometries and converting them into signed distance fields (SDFs), to orchestrating high-performance computing (HPC) simulations, and finally, preparing the simulation results into ML-ready, multi-resolution data. This configuration-driven pipeline ensures reproducibility, enables systematic studies, and helps researchers build better AI models for fluid dynamics by providing consistent, scalable datasets.
The field of computational fluid dynamics (CFD) is increasingly leveraging machine learning (ML) to predict flow behavior and optimize designs. However, a significant hurdle has been the lack of standardized, ML-ready datasets and reproducible pipelines to create them. Researchers often rely on custom, one-off preprocessing steps, making it difficult to compare studies, conduct scaling analyses, or perform systematic experiments.
Addressing this critical gap, a new framework called ChannelFlow-Tools has been introduced. This configuration-driven pipeline aims to standardize the entire process of generating datasets for 3D obstructed channel flows, from initial geometry creation to producing ML-ready inputs and targets. The core idea is to transform bespoke data generation into a repeatable, configurable workflow for CFD surrogate modeling.
A Comprehensive Data Generation Pipeline
ChannelFlow-Tools integrates several key stages, all governed by a single configuration file, ensuring end-to-end consistency and reproducibility:
1. Geometry Generation: The pipeline begins by programmatically synthesizing 3D obstacle geometries within a channel. It uses CadQuery to create single or multi-object obstacles from six primitive families (cube, cone, cylinder, sphere, torus, wedge). Crucially, it includes feasibility checks to ensure objects are in-bounds, do not intersect, and meet minimum volume requirements. Each accepted scene is exported as a triangulated STL mesh along with detailed metadata, capturing all parameters for exact reproduction.
2. Signed Distance Field (SDF) Generation: The generated STL meshes are then converted into Signed Distance Fields (SDFs). An SDF represents the distance from any point in space to the surface of the obstacle, with a negative sign inside the object and a positive sign outside. This stage uses OpenVDB for sparse level-set representation and then rasterizes it into dense NumPy arrays, which are ideal for machine learning models. The process allows for explicit control over voxel spacing and narrow-band width, and supports various resolutions (e.g., 128x32x32, 256x64x64, 512x128x128).
3. Solver Automation: For simulating fluid flow, ChannelFlow-Tools provides an automation layer that orchestrates solver campaigns on high-performance computing (HPC) systems. It stages all necessary inputs, generates a deterministic run manifest, and launches/monitors batches of simulations. The simulations are executed using the waLBerla lattice–Boltzmann method (LBM) framework, configured for 3D obstructed channel flows. Solver parameters, like domain size and boundary conditions, are derived from the same central configuration, maintaining consistency across the pipeline.
4. Resampling and Down-sampling: Finally, the simulation outputs are converted into co-registered Cartesian tensors at multiple resolutions, ready for ML ingestion. This stage uses a ParaView/VTK-based resampling path, allowing for transparent kernel and footprint choices. This ensures that the data is consistently formatted and aligned with the SDF inputs, facilitating fair comparisons and scaling studies for ML models.
Validating the Dataset’s Utility
As a case study, the researchers generated approximately 10,000 scenes spanning a wide range of Reynolds numbers (Re≈100 –15,000) with diverse obstacle shapes and poses. To demonstrate the practical utility of the generated datasets, a minimal 3D U-Net model was trained on progressively larger subsets of this data. The results showed a monotonic reduction in RMSE (Root Mean Squared Error) and MAE (Mean Absolute Error) as the dataset size increased. This predictable scaling behavior illustrates that the standardized representations produced by ChannelFlow-Tools effectively support reliable machine learning studies.
Also Read:
- Automating 3D Design: GenCAD-3D Generates Editable CAD Programs from Scans
- RLinf: Smarter Scheduling for Faster Reinforcement Learning
Impact and Future Directions
ChannelFlow-Tools represents a significant step towards standardizing the geometry-to-simulation-to-ML export path for obstructed channel flows. By providing an open, configuration-first pipeline and a feature-complete reference dataset, it aims to reduce setup friction, improve comparability across studies, and enable systematic scaling and ablation work in CFD surrogate modeling. This framework helps researchers move away from custom, project-specific data preparation, fostering a more collaborative and reproducible environment for AI in engineering domains.
For more in-depth information, you can read the full research paper here.
