AI-Assisted Floor Plan Design Incorporating Structural Constraints

Abstract

Automatic layout generation studies remain limited in their ability to ensure structural continuity in floor plans. In response, this study considers structural schema as a design constraint. Generative Adversarial Network algorithms for image-to-image translation were utilised to generate floor plans. The dataset, consisting of 552 housing floor plans with structural layouts, was labelled and trained to generate house plan alternatives to assist the interior layout design process in early design phases. The dataset was trained with the machine-learning algorithms Pix2Pix and BicycleGAN. The comparative evaluation of the results using Learned Perceptual Image Patch Similarity indicates that BicycleGAN performs better than Pix2Pix, and the suggested workflow is quite promising. However, the lack of circulation areas was identified as a common limitation in both models. This workflow also has the potential to be used for renovation purposes.