Researchers have developed ChannelFlow-Tools, an open-source pipeline for generating datasets of three-dimensional obstructed channel flows. The pipeline allows for controlled data generation and can regenerate or adapt existing datasets to match specific research requirements. It integrates various stages, including procedural obstacle geometry generation, lattice-Boltzmann simulation, and packaging into machine-learning-ready tensors. The workflow is driven by configuration files, ensuring reproducibility. ChannelFlow-Tools has been evaluated through various tests, including mesh-integrity audits and validation of the signed-distance-field representation. Three surrogate models were trained on a sample dataset generated entirely through the pipeline, demonstrating physically consistent and directly usable training data. --- Why it matters: This matters to engineers and researchers in AI because it provides a shared infrastructure for controlled benchmarking of geometry-aware computational fluid dynamics (CFD) surrogates. This can help improve the reliability of CFD models by providing high-quality training data. Source: https://arxiv.org/abs/2509.15236