EVALUATING WINDOW-TO-WALL RATIO IN GENERATIVE AI ARCHITECTURAL DESIGN: INSIGHTS FROM SHAP ANALYSIS AND PREDICTIVE MODELING

This study evaluates the performance of Stable Diffusion in generating window-to-wall ratios (WWRs) in architectural design, focusing on aligning generative AI outputs to performance-orient standards. By employing a combination of predictive modeling using Support Vector Regression (SVR) and Shapley Additive Explanations (SHAP) analysis, this research investigates the interplay between prompt parameters and their impacts on the generated WWRs. The study identifies key prompt parameters such as building orientations, floor heights, and window designs as crucial parameters in shaping the generated WWRs, revealing the potential for targeted prompt engineering to optimize AI-generated architectural designs. The findings show that although Stable Diffusion well captures complicated prompt semantics, it shows limitations in generating ideal WWR targets and accommodating location-specific needs. The results highlight the need to incorporate performance-driven strategies into generative AI processes to close the gap between architectural performance and generative AI design.