ChannelFlow-Tools: Machine-Learning-Ready Datasets of 3D Obstructed Channel Flows

ChannelFlow-Tools: Machine-Learning-Ready Datasets of 3D Obstructed Channel Flows This data release accompanies the ChannelFlow-Tools paper. It contains four related datasets supporting the generation, verification, and machine learning experiments reported in the manuscript. Each dataset is provided as a separate ZIP archive. Release_v1_256_128_128_domain.zipFlow dataset on the 256 × 128 × 128 grid with Δx=4 lu. It contains 450 complete simulation scenes. For each scene, the release provides the signed distance field, time averaged velocity field, obstacle force log, source STL geometry, Reynolds number, and associated metadata. The geometry and flow data are co registered on a common grid. The archive also includes the HDF5 dataset, standalone STL and SDF data, the per scene manifest, and the train, validation, and test splits. Release_v2_128_64_64_domain.zipFlow dataset on the 128 × 64 × 64 grid with Δx=8 lu. It contains the same 450 simulation scenes as the higher resolution release, represented on the coarser ML grid. Release_geometry_stl_corpus.zipProduction geometry corpus containing 10,060 scene level STL meshes, representing 15,562 individual objects, with corresponding YAML metadata. These geometries are used in the geometry generation validation reported in the paper. The archive also includes the Sobol sampling schedule and generation metadata. Release_sdf_corpus.zipSDF verification subset containing 350 geometry–SDF pairs covering all six single object shape families as well as two object and three object configurations. The corpus scale SDF validation reported in the paper is performed on the 256 × 128 × 128 grid with Δx=4 lu. The archive provides the corresponding data required to reproduce and inspect this validation. Each ZIP archive contains its own README, DATASHEET, and per scene manifest with SHA-256 checksums. The two flow datasets additionally contain the train, validation, and test split definitions. The top level README.md describes the four datasets and their relationships. The release covers six obstacle shape families: sphere, cuboid, cylinder, cone, torus, and wedge. Scenes contain between one and three obstacles, covering all 83 unordered shape compositions. The computational domain is 1024 × 512 × 512 lattice units. The SDF sign convention is negative inside the obstacle and positive in the fluid region. Dataset specific preprocessing and coordinate information are documented in the corresponding README and DATASHEET files. Licence: CC BY 4.0 Please cite the accompanying ChannelFlow-Tools paper when using this dataset. Citation information is provided in CITATION.cff.