AirfoilMNIST: A Large-Scale Dataset based on Two-Dimensional RANS Simulations of Airfoils

Studying aerodynamic problems is still dominated by computationally heavy simulations which require substantial resources and knowledge about aerodynamics and numerical methods. While first adaptions of artificial intelligence have reached the field of aerodynamics to accelerate this process, data is still scarce and often lacks the proper validation. To bridge this gap, we propose airfoilMNIST, a comprehensive dataset of two-dimensional, RANS-based flow fields for NACA 4- and 5-series airfoils. Our dataset consists of around 150’000 samples for Mach numbers up to 0.6 and angles of attack −5 &lt;= alpha &lt;= 15 where each sample contains the mean flow fields for density, eddy viscosity, pressure, temperature, turbulent kinetic energy, turbulent thermal diffusivity, specific rate of dissipation and velocity. <br> The dataset is separated into three subsets: - airfoilMNIST-raw: Individual CFD samples which are stored in the VTU file format together with all airfoil geometries as STL files. Additionally, each sample has an associated text file with the progression of the force and moment coefficients for the respective iteration number of the CFD simulation. The entire subset has a total size of 3.5TB - airfoilMNIST: Preprocessed dataset compatible with the Tensorflow Data API. Flow fields have been resampled onto a uniform grid and stored as two NumPy arrays, designed for encoder-decoder architectures. Encoder array is a field with the initial and boundary conditions while the decoder array are the individual flow fields. Dataset has a predefined, randomised train-test split of (80, 20)% and a total size of 500GB. - airfoilMNIST-incompressible: Further simplification of the airfoilMNIST subset. Only incompressible simulations (Mach &lt; 0.3) are stored. Instead of storing all flow fields, only the pressure and velocity fields normalised with the freestream conditions are saved. Dataset has a predefined, randomised train-test split of (80, 20)% and a total size of 50GB.

当前气动问题的研究仍以计算量极大的数值模拟为主，这类模拟需要耗费大量计算资源，同时要求研究者具备扎实的气动学与数值方法相关知识。尽管人工智能的初步应用已进入气动领域以加速这一流程，但相关数据集仍较为匮乏，且大多缺乏有效的验证。

为填补这一空白，我们提出airfoilMNIST数据集——一款针对NACA 4系列与5系列翼型的二维、基于雷诺平均Navier-Stokes（Reynolds-Averaged Navier-Stokes, RANS）流场的综合数据集。本数据集包含约15万个样本，覆盖马赫数不超过0.6、攻角α范围为-5°≤α≤15°的工况，每个样本均包含密度、涡粘性、压强、温度、湍流动能、湍流热扩散率、比耗散率以及速度的平均流场数据。

本数据集分为三个子数据集：
- airfoilMNIST-raw：原始计算流体动力学（Computational Fluid Dynamics, CFD）样本集，所有翼型几何均以STL格式存储，单个CFD样本则以VTU格式保存。此外，每个样本均附带关联文本文件，记录对应CFD模拟迭代步数下的力与力矩系数变化情况。该子数据集总大小为3.5TB。
- airfoilMNIST：经过预处理的数据集，兼容TensorFlow数据API。流场已被重采样至均匀网格，并以两个NumPy数组的形式存储，专为编码器-解码器架构设计。其中编码器数组包含初始条件与边界条件场，解码器数组则对应各单个流场。该数据集已预先完成80%训练集、20%测试集的随机划分，总大小为500GB。
- airfoilMNIST-incompressible：airfoilMNIST子集的进一步简化版本，仅存储马赫数小于0.3的不可压缩模拟数据。相较于完整流场存储，该子集仅保存以自由来流条件归一化后的压强与速度场。该数据集同样预先完成80%训练集、20%测试集的随机划分，总大小为50GB。