Cashen/optiwing3d_engibench

OptiWing3D (Structured Parquet)

Structured Parquet version of OptiWing3D, prepared for use with the Hugging Face datasets library.

Splits

• Split by case_num (no slice-level leakage)
• 70 / 15 / 15 (train / validation / test)
• SEED = 0
• Deterministic and reproducible

All rows belonging to the same case_num are assigned to the same split.

Directory Structure

optiwing3d_hf/ └── optiwing3d_structured/ └── data/ └──test_cases.csv └──test.parquet └──train_cases.csv └──train.parquet └──validation_cases.csv └──validation.parquet ├── make_splits_parquet.py ├── README.md

Data Format

• Stored in Apache Parquet
• Numpy array columns converted to nested Python lists
• Each row corresponds to a single 3D wing slice/sample

Array-based columns:

• coords
• transforms
• te_shifts
• coef_pressure
• velocity_x
• velocity_y
• velocity_z

Loading

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={
        "train": "optiwing3d_structured/data/train.parquet",
        "validation": "optiwing3d_structured/data/validation.parquet",
        "test": "optiwing3d_structured/data/test.parquet",
    },
)

To regenerate the splits from df_raw_3d.pkl:
    python make_splits_parquet.py