Space Layout Optimization For Natural Ventilation Using Machine Learning Techniques

Abstract

Indoor airflow distribution holds significant importance in evaluating the natural ventilation conditions of buildings. Conventionally, Computational Fluid Dynamics (CFD) is used for evaluating airflow patterns in a multizone space layout during the early design phase. While CFD can provide accurate airflow information, its adoption requires substantial computational resources and considerable running time, which limits its application in the fast pace of early-stage architectural design. With the recent rapid advancements in machine learning techniques, data-driven surrogate models have emerged as a potential alternative to CFD solvers for fast prediction of flow fields. While delivering promising performance on generic boundary geometries, machine learning approaches on multizone indoor airflow are less explored due to its inherent complexity and limited availability of high-quality datasets. This dissertation introduces a new machine learning framework designed for the fast prediction of CFD-generated airflow fields in multizone space layouts and the optimization of space layouts based on airflow distribution. The framework consists of three components: a data generation module, an ensemble learning framework, and a model predictive optimizer. The data generation module randomly generates multizone space layouts and populates their 3D flow patterns using conventional CFD techniques. The ensemble learning framework leverages multiple machine learning models to collectively predict multizone airflow fields, surpassing standard machine learning approaches in terms of accuracy. The framework also demonstrates robust functionality by incorporating physics-informed loss functions and exhibits adaptability for predicting more complex scenarios beyond the training dataset. The optimizer uses the framework as a predictive model to minimize the portion of undesired indoor air velocity with optimal window positions. Furthermore, a design tool is implemented using the methods presented in this research to facilitate early-stage natural ventilation design.