Using Diffusion Models for Reducing Spatiotemporal Errors of Deep Learning Based Urban Microclimate Predictions at Post-Processing Stage

Abstract

Computational fluid dynamics (CFD) is a powerful tool for modeling turbulent flow and is commonly used for urban microclimate simulations. However, traditional CFD methods are computationally intensive, requiring substantial hardware resources for high-fidelity simulations. Deep learning (DL) models are becoming popular as efficient alternatives as they require less computational resources to model complex non-linear interactions in fluid flow simulations. A major drawback of DL models is that they are prone to error accumulation in long-term temporal predictions, often compromising their accuracy and reliability. To address this shortcoming, this study investigates the use of a denoising diffusion probabilistic model (DDPM) as a novel post-processing technique to mitigate error propagation in DL models' sequential predictions. To address this, we employ convolutional autoencoder (CAE) and U-Net architectures to predict airflow dynamics around a cubic structure. The DDPM is then applied to the models' predictions, refining the reconstructed flow fields to better align with high-fidelity statistical results obtained from large-eddy simulations. Results demonstrate that, although deep learning models provide significant computational advantages over traditional numerical solvers, they are susceptible to error accumulation in sequential predictions; however, utilizing DDPM as a post-processing step enhances the accuracy of DL models by up to 65% while maintaining a 3 times speedup compared to traditional numerical solvers. These findings highlight the potential of integrating denoising diffusion probabilistic models as a transformative approach to improving the reliability and accuracy of deep learning-based urban microclimate simulations, paving the way for more efficient and scalable fluid dynamics modeling.