Dataset and Surrogate models based on 2D RANS thermal street canyon pollutant dispersion.

Dataset Description This dataset was generated using Reynolds-Averaged Navier–Stokes (RANS) simulations to model pollutant dispersion in a 2D idealized street canyon. It contains 49,590 samples, divided into training+validation and test subsets. Each sample represents a 64×64 pollutant concentration field, produced as part of a data augmentation study for surrogate modeling. Data Structure Concentration fields are stored in NumPy .npy format and can be loaded using numpy.load(file).The arrays have shape: (n_samples, 64, 64) or (n_samples, 4096), which are interconvertible via .reshape(). Input parameters associated with each sample are stored in a separate .npy file and correspond positionally (i.e., the i-th row matches the i-th sample in the concentration dataset). The mask file (in DataCoordinates/Mask.npy) identifies grid points inside building areas, which should be excluded from statistical analyses or visualizations of the street canyon. Input Parameters Each sample is defined by 18 input parameters representing physical, turbulence, and emission characteristics. The values are normalized, and can be restored using the inverse_norm_parameters function provided in the code base. The parameters and their physical meaning are: Sct – Turbulent Schmidt number Ce1 – k−ε model parameter Ce2 – k−ε model parameter Cμ – k−εk model parameter σk – k−εk model parameter B – k−ε model parameter κ – von Kármán constant Uτ – Friction velocity y0 – Aerodynamic roughness length pk – TKE source scaling factor pϵ – Dissipation rate source scaling factor Bk – Background concentration offset pB – Background concentration slope Q – Emission rate Qh – Emission source height Qp – Heat release rate θ – Solar incidence angle ΔT – Background temperature increase Surrogate Models Two surrogate modeling approaches based on deep learning are trained and evaluated using this dataset: PCA + MLP: A two-step model combining Principal Component Analysis (PCA) for dimensionality reduction and a Multi-Layer Perceptron (MLP) to predict the modal coefficients. U-Net: A convolutional encoder–decoder architecture designed for structured image prediction. Both models are optimized using the NSGA-II algorithm and included in the TrainedModels/ directory. Usage To demonstrate how to load and evaluate the models using the dataset, we provide two example scripts: MLPExample.py – loads the MLP surrogate, reconstructs PCA-based fields, and plot a sample. UnetExample.py – loads the Unet surrogate and plot a sample. Each model requires loading associated parameter and coordinate data, as well as the building mask. For more details on the simulation setup, optimization, and evaluation strategy, please refer to the associated preprint Assessment of deep-learning strategies for surrogate modeling of pollution dispersion in a thermal street canyon, Beltrami, G., et al. 2025, submitted to computer & fluids. The authors acknowledge financial support from the AIR-URBAN project (TED2021-130210A-I00/ AEI/10.13039/501100011033/ European Union NextGenerationEU/PRTR).