Combining LES, Machine Learning, and Reduced-Order Models for Predicting Natural Ventilation

Energy use for building cooling is projected to increase dramatically in the coming decades, driven by urbanization and climate change. In response, passive strategies, such as natural ventilation and cooling, provide energy-efficient and sustainable alternatives. The effectiveness of these approaches is highly sensitive to local airflow patterns within the urban canopy and building interiors. Accurately resolving these geometry-specific airflow patterns requires computational fluid dynamics, and ideally large eddy simulations (LES) that resolve the large-scale separation regions and wakes that are typical in urban flows. When resolving the coupled outdoor-indoor flow, these simulations become computationally intensive.

The objective of this work is to establish a framework that combines physics-based simulations, machine learning, and reduced-order ventilation models to accurately predict natural ventilation performance at low computational cost. Previous work has shown that exterior pressure distributions are strong predictors of the ventilation rates, indicating that a detailed solution of the interior flow is not essential. Furthermore, machine learning models trained on LES data present new opportunities for predicting urban canopy flows at reduced computational cost. In this work, we apply a U-Net model trained on large-eddy simulation data to predict surface pressure fields within realistic urban canopies. These predicted pressures are then coupled with a 1-D model to estimate natural ventilation rates in specific building geometries. The proposed approach can support the accurate assessment of natural ventilation and cooling across broad geographic regions.