Request for exact AhmedML and DrivAerML train/validation/test splits

Hi @hangzhou188 and @wuhaixu2016,

Thank you for releasing Transolver-3. I am working to reproduce the AhmedML and DrivAerML experiments from the paper.

The paper describes random 400/50/50 train/validation/test partitions for both datasets, but the exact case identifiers are not specified. The repository currently includes drivaerml.json with 400 training runs and 50 test runs, but it does not list the validation cases, and I could not find a corresponding AhmedML split.

Could you please share the exact train, validation, and test case IDs used for the reported results on both AhmedML and DrivAerML? It would also be helpful if you could confirm that the checked-in drivaerml.json matches the split used in the paper and clarify which DrivAerML validation cases were used, including the 16 hidden cases without simulation results mentioned in Appendix A.1.

A JSON file, split-generation script, or random seed would be greatly appreciated.

Thank you!

Start by reading drivaerml.json and checking the repository for AhmedML and DrivAerML split artifacts. Compare the checked-in cases with the paper's 400/50/50 description. Done means documenting or adding the exact train, validation, and test IDs, the generation seed or script, and the status of the 16 hidden DrivAerML validation cases.