Refined ISO model for wind-driven rain (WDR) on buildings: Machine learning-based approach

• Published in: Building and environment Vol. 285; p. 113522
• Main Authors: Gholamalipour, Payam, Stathopoulos, Ted, Ge, Hua
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2025
• Subjects: Computational fluid dynamics (CFD); ISO semi-empirical method; Machine learning (Ml); Wall Factor, W [-]; Wind-driven rain (WDR); Wall Factor, W [-]; Machine learning (Ml); Wind-driven rain (WDR); ISO semi-empirical method; Computational fluid dynamics (CFD)
• ISSN: 0360-1323
• DOI: 10.1016/j.buildenv.2025.113522

•ISO model shows up to 2–4 times overpredicting Wind-Driven Rain loading on buildings.•Some Machine Learning models (e.g., ANN) outperform ISO in real-world WDR predictions.•ANN model reduces the average error of refined ISO Wall Factor by up to 53 %. Wind-Driven Rain (WDR) is a main source of moisture impacting building facade, which can negatively influence their hygrothermal performance and long-term durability. The WDR deposition patterns on building facades vary significantly based on different meteorological and geometrical parameters. Among the three main approaches for studying WDR on buildings (experimental, numerical, and semi-empirical), the ISO semi-empirical method is widely used despite being subject to substantial errors under certain situations (2–4 times overpredictions). A major source of these errors is the suggested constant Wall Factor, which is limited to six primary building configurations by ISO semi-empirical model. This study aims to improve the ISO semi-empirical model by generating refined Wall Factors for any flat roof stand-alone building configurations using ML approach through three main steps. Part A, using Computational Fluid Dynamics (CFD) to quantify WDR loading for various meteorological and geometrical parameters and generate a comprehensive dataset. The steady-state standard k-ω RANS turbulence model is coupled with Eulerian Multiphase (EM) WDR techniques (referred to as RANS-EM) to simulate WDR. This approach has been validated with existing wind-tunnel and field measurement data. Part B uses the CFD-generated Wall Factor dataset to train various Machine Learning (ML) models (i.e., Decision Trees Regression (DTR),), and Deep Learning (DL) model (i.e., Artificial Neural Network (ANN)), aiming to generate refined Wall Factors applicable across a broad range of parameters. Part C compares the WDR estimation using the ISO Wall Factor and the refined Wall Factor against field measurement results, demonstrating an average error reduction of 53 %.