Optimization of tall buildings subjected to wind load using genetic algorithm and image-based machine learning

Abstract

As the prevalence of tall buildings are on an upward trend in urban cities, the need to navigate their design intricacies becomes increasingly important. Tall buildings exhibit dynamic and nonlinear responses to applied load and are required to satisfy extensive design requirements, leading to the need for complex analysis techniques to evaluate very specific engineering problems. Technological advances in supporting research fields provide engineers with both computational resources and algorithms for use as tools not previously available, strengthening the case for Machine Learning (ML) and surrogate modelling techniques to assist with the interpretation and exploration of design spaces in advanced analysis. This thesis studies the use of Convolutional Neural Networks (CNNs) designed to simultaneously facilitate the optimization of Reinforced Concrete (RC) shear wall topology and size. As an image-based approach, the work assesses the capability of the proposed algorithms to generalize the abstract relationship between structural layout and numerical performance metrics of tall building designs. The resulting models display significant capability of replicating Finite Element Analysis (FEA) corresponding to structural layout images. Further, an optimization framework is developed which utilizes these models and a Genetic Algorithm (GA) to automate the design processes which typically rely on the expertise of engineers. As a result, both architects and engineers can utilize the proposed framework to identify design solutions corresponding to a reduced quantity of shear wall volume, and/or total number of piers required, achieving cost savings in real-world applications. This work reveals the potential of CNN-based surrogate models in the design of tall buildings, especially when proposed for structural and multidisciplinary optimization. KEYWORDS Tall buildings, wind load, shear wall, Reinforced Concrete (RC), Latin Hypercube Sampling, Convolutional Neural Network, Genetic Algorithm (GA), multi-objective optimization, wind engineering, surrogate model, deep learning, performance-based design, Computational Fluid Dynamics (CFD), structural optimization, Design Space Exploration (DSE), Finite Element Analysis (FEA).