ChannelFlow-Tools: A Configuration-Driven Pipeline for Generating Machine-Learning-Ready Datasets of 3D Obstructed Channel Flows

ChannelFlow-Tools is a new open-source pipeline for building machine-learning-ready datasets of 3D obstructed channel flows, with an emphasis on reproducibility and controlled data generation. Instead of relying on one-off precomputed datasets, the workflow is driven by configuration files so geometry, simulation settings, and exports can be regenerated or extended for new surrogate-model studies.

The pipeline covers the full path from procedural obstacle generation to signed-distance-field voxelization, lattice-Boltzmann simulation, and packaging into tensors that are ready for training. Six obstacle shape families are supported, and byte-identical reproducibility was verified for the geometry-generation stage. The dataset also went through mesh-integrity checks, SDF validation, solver benchmarks using canonical sphere-flow cases, and per-scene data-integrity audits.

To prove the output is practical for downstream ML, the team trained three surrogate models — 3D U-Net, FNO, and U-FNO — on 450 simulations spanning roughly Re_c 1000 to 10,000. Those models learned the geometry-to-flow mapping and behaved sensibly on out-of-distribution tests across both shape family and Reynolds number, which is the real signal here: the data is not just clean, it is useful for benchmarking geometry-aware CFD surrogates.

For game developers, the immediate relevance is less about CFD itself and more about the pipeline philosophy. If you are building simulation-heavy tools, procedural content systems, or ML-assisted authoring workflows, this is a strong example of how to make synthetic...