sunweiwei/ai4sci-hydrogym-research-v2

HydroGym Research v2: actual CFD resources

This repository hosts data, not the Harbor task or training/evaluation code. Harbor instructions, starter/reference implementations, model-delivery contract and evaluator belong in T0-RSI/ai4sci-tasks.

Publication status

The actual resource upload is complete: all 558 selected files and their exact 453,756,315,982-byte total are present. Use the immutable Hub revision pinned in the GitHub task's environment/data/download-source.json, not a floating main revision. The earlier manifest/code-only publication was a packaging mistake and is not the data release; its Harbor files have been removed from the current Hub tree.

Contents

environments/CASE/... contains the original geometry, grids, boundary/solver configurations and initial flow fields for the selected 39 tasks (20 2D, 19 3D): 558 files, 453,756,315,982 bytes. release.json records the frozen selection and profiles. No pretrained controller or trajectory archive is added. Training and evaluation share these source assets; RL experience is generated through interaction with the simulator.

These bytes are an unchanged subset of the authors' public dataset:

• —Source: dynamicslab/HydroGym-environments.
• —Source commit: cf5ef6748443ae72e9647f398487e98d421a7243.
• —Release ID: hydrogym-public-research-v2.
• —Canonical release hash: cabc8b718cdf9b22fcd661b9b3c5f8e452498e1c299b0fc9c7ac1afc22df47d2.

Credit for the original HydroGym resources belongs to their authors. This mirror does not change source ownership or upstream asset/solver terms. The upstream data card at the pinned revision does not declare a dataset license; no new blanket license is asserted here. Solver binaries are not redistributed here.

How to use

Use the materializer in the GitHub Harbor task, which downloads the selected files from this Hub repository at its pinned revision, checks sizes and Git/LFS hashes, and writes the local materialization receipt. Do not copy a receipt from another machine. Bind the prepared directory read-only at /data for training and independent evaluation. The 454 GB source budget excludes Docker images, simulation scratch, generated experience and model checkpoints.