Assessment of deep-learning strategies for surrogate modeling of pollution dispersion in a thermal street canyon

Data-driven surrogate models provide an effective means to reduce the high computational cost of computational fluid dynamics (CFD), enabling large-scale design and uncertainty studies that would otherwise be infeasible. Most existing surrogate models for CFD of pollution dispersion focus primarily on variations in boundary conditions, while other uncertain factors are rarely considered. We extend the input space beyond boundary conditions to include both modeling uncertainties (turbulence parameters) and key physical drivers such as emission strength, source characteristics, inlet profiles, and solar irradiance, creating a database with nearly 50000 converged samples. We emulate the two-dimensional RANS simulations of the canonical thermal street-canyon problem in a simplified grid of 64 × 64 points within the canyon. We benchmark two surrogate modelling approaches, namely Principal Component Analysis combined with a multilayer perceptron (PCA+MLP) and a Dense layer Convolutional Encoder-Decoder with skipped connections (DCED) surrogate model to contrast linear and nonlinear dimensionality reduction strategies. To assess performance, we introduce a combined regression–classification error framework based on vortex-direction prediction. PCA+MLP is computationally efficient, while DCED delivers slightly higher accuracy at larger training and inference costs. The PCA+MLP model attains a regression error of 4.42% and a classification error of 2.24%, while providing an inference speed-up of approximately five orders of magnitude relative to a single iteration of the CFD model. The DCED achieves slightly lower errors (3.89% regression and 2.20% classification), with an inference acceleration of about three orders of magnitude. For both models, performance decreases in thermally dominated regimes, highlighting the challenge of capturing strongly nonlinear flow phenomena. Our results also emphasize the importance of dataset balance and suggest targeted sampling strategies for better generalization. We publicly release our CFD benchmark database to support reproducibility and future research.