Evaluating daylight performance of AI-generated housing plans via diffusion models and climate-based simulation

Abstract This study investigates the applicability of generative artificial intelligence (AI) in early-stage architectural design by evaluating the daylight performance of AI-generated sustainable housing plans across five distinct climate zones. A three-phase methodology was implemented: (1) Plan generation using text-to-image diffusion models (ChatGPT, Copilot, and LookX); (2) digital reconstruction in AutoCAD; and (3) daylight simulation via Velux Daylight Visualizer. Climate-adaptive prompts were formulated to guide the AI tools in producing context-specific floor plans with passive strategies. Out of 31 initial plans, eight valid outputs (five from ChatGPT and three from Copilot) were reconstructed in AutoCAD and simulated. Quantitative simulations were conducted on equinox and solstice dates, and average illuminance values were analyzed for key interior spaces (living room, kitchen, and bedroom). ChatGPT-generated plans demonstrated higher spatial clarity and more balanced daylight performance, whereas Copilot outputs varied significantly, and LookX was excluded due to insufficient architectural legibility. Results revealed that none of the models consistently integrated solar orientation or seasonal lighting considerations, indicating a gap between generative representation and environmental logic. The research contributes a replicable workflow that bridges generative AI and performance-based evaluation, offering critical insight into the current limitations and future potential of AI-assisted architectural design. The findings underscore the need for next-generation AI systems capable of semantic, spatial, and climatic reasoning to support environmentally responsive design practices.