Artificial Intelligence-Powered Computational Strategies in Selecting and Augmenting Data for Early Design of Tall Buildings with Outer Diagrids

Abstract

In the realm of architectural computing, this study explores the integration of parametric design with machine learning algorithms to advance the early design phase of tall buildings with outer diagrid systems. The success of such an endeavor relies heavily on a data-driven and artificial intelligence-enhanced workflow aimed at identifying key architectural and structural variables through a feature/response selection process within a supervised machine learning framework. By augmenting an initial dataset, which was notably limited, through four distinct techniques—namely Gaussian copula, conditional generative adversarial networks, Gaussian copula generative adversarial network, and variational autoencoder—this study demonstrates a methodical approach to data enhancement in architectural design. The results indicate a slight preference for the Gaussian copula method, attributed to its less complex hyperparameter tuning process. Evaluation through a random forest regressor revealed stable performance across various cross-validation techniques on synthetic data, although with an acceptable decrease in the coefficient of determination, from an original average score of 0.925 to an augmented score of 0.764. This investigation underscores the potential of artificial intelligence-powered computational tools to guide design decisions by pinpointing the variables with the most significant impact on relevant outputs, quantitatively assessing their influence through the accuracy of the employed machine learning methods.

Keywords

• AI-enabled parametric architectural design
• supervised machine learning
• architectural form generation
• feature selection
• data augmentation
• regression