Prediction and optimization of daylight performance of AI-generated residential floor plans

Abstract

The integration of artificial intelligence (AI) into architectural design, particularly in generating floor plans, has the potential to significantly streamline the design process. However, most AI-generated floor plans focus primarily on form and spatial arrangement, often neglecting essential performance evaluations due to their output being rendered as images without necessary geometries and properties for effective physical modeling analysis. To address this limitation, we propose a novel methodology that combines diffusion models with Generative Adversarial Networks (GANs) to simultaneously generate and evaluate architectural floor plans. We fine-tuned a Low-Rank Adaptation (LoRA) model on a dataset to create residential floor plans, while the GAN model facilitates rapid predictions of daylight performance. Our results demonstrate that the diffusion model can generate a variety of floor plans, exceeding the types included in the training set. The GAN model provided accurate assessments of daylight performance by deviating no more than 5 % from ground truth, achieving a mean squared error (MSE) as low as 4.2, and a structural similarity index (SSIM) reaching 0.98 on the test set. The proposed workflow operates 267 times faster than traditional modeling-simulation workflows. This integrated approach equips architects with an efficient and reliable tool for early-stage design decisions, enhancing the effectiveness of AI-driven architectural workflows.