Kokoslocke/NACA_4_Digit_for_ML

NACA 4-Digit Airfoil CFD Dataset

Point-cloud CFD solutions for NACA 4-digit airfoils, generated with OpenFOAM v13 (k-ω SST). Intended for training surrogate models that predict steady-state flow fields from airfoil geometry and flow conditions.

Dataset Summary

• —~850 converged in-distribution cases across 50 distinct NACA 4-digit profiles
• —AoA range: −5° to +5°
• —Reynolds number range: 100,000 – 500,000
• —129 out-of-distribution (OOD) probe cases at high Re (1–2 × 10⁶) and high AoA (10–15°) for generalization evaluation
• —Split strategy: profile-level — all cases for a given airfoil shape belong to a single split, preventing geometry leakage between train/val/test

Airfoil Parameter Space

File Format

Each case is stored as a compressed NumPy archive (.npz). The bounding box is chord-normalized with the leading edge at x = 0 and the trailing edge at x = 1.

Spatial domain (bounding box):

x ∈ [−1.5, 3.5]   (1.5c upstream, 2.5c downstream)
y ∈ [−1.5, 1.5]   (1.5c above and below)

Arrays per .npz file

N varies per case (typically 50,000 – 150,000 cells inside the bounding box).

Airfoil-surface arrays

In addition to the volume point cloud, each file carries a separate surface table with one row per airfoil wall face (M rows, M ≈ N_wall, typically a few hundred). These live on the airfoil surface itself, unlike is_wall which marks the first layer of volume cells sitting just off the wall. Rows are ordered by wall-face index and are mutually aligned.

Availability: the surface arrays are present for all splits — train, val, test, and ood.

Quantities are kinematic (divided by density, consistent with p), matching OpenFOAM's wallShearStress function object. Multiply by ρ for physical units. Skin friction and pressure coefficients follow directly:

import numpy as np

data = np.load("NACA2412_p3.5_2.0e5.npz")
u_mag = float(np.hypot(data["u_init"][0], data["v_init"][0]))
q = 0.5 * u_mag**2                      # kinematic dynamic pressure
cf = np.linalg.norm(data["wall_shear"], axis=1) / q
cp = data["wall_p"] / q
# Chordwise wall-shear sign flags separation (τ_w,x < 0 → reversed flow):
tau_x = data["wall_shear"][:, 0]

Loading a sample

import numpy as np

data = np.load("NACA2412_p3.5_2.0e5.npz")
x, y = data["x"], data["y"]
u, v, p = data["u"], data["v"], data["p"]
re = float(data["reynolds"])

File Naming Convention

NACA{code}_{sign}{aoa}_{Re}.npz

• —{code} — 4-digit NACA identifier (e.g., 2412)
• —{sign} — p for positive AoA, n for negative AoA
• —{aoa} — angle of attack in degrees (one decimal place)
• —{Re} — Reynolds number in scientific notation (e.g., 2.0e5)

Example: NACA2412_p3.5_2.0e5.npz → NACA 2412 airfoil, AoA = +3.5°, Re = 200,000.

Dataset Structure

NACA_4_digit_for_ml/
├── README.md
├── metadata.csv           # Per-case tabular summary (see below)
└── *.npz                  # One file per converged case

metadata.csv columns

Dataset Splits

Splits are profile-level: every case sharing an airfoil code is assigned to one partition only. This ensures the model cannot memorize geometry.

The split field is also stored in each case's meta.yaml in the source repository. The OOD cases use airfoil shapes not present in train/val/test (95 unique new geometries, including thickness 6–26 % outside the 8–18 % training range) at Reynolds numbers and angles of attack well outside the training envelope (Re 1–2 × 10⁶, |AoA| 10–15°).

CFD Setup

Only cases that converged within the iteration budget are included. Non-converged cases are logged in the source repository's dataset/rejection_log.csv.

Source

Generated with cfd_data_generator. Mesh and field data produced by the pipeline in dataset/scripts/; ML-ready .npz files assembled by dataset/scripts/build_ml_dataset.py.