Automated door placement in architectural plans through combined deep-learning networks of ResNet-50 and Pix2Pix-GAN

Abstract

Previous studies on automating building design with deep learning primarily focused on planning room layouts, limiting the design of architectural elements such as doors and windows. This led to the misalignment of rooms, which in turn resulted in a design proposal with no space for architectural openings. In particular, the placement of a door that can set the circulation throughout the building and the privacy gradient as a connection between rooms is still determined by a rule or a designer. To overcome these limitations, this study conducted automated door placement in architectural plans through combined deep-learning networks of ResNet-50 and Pix2Pix-GAN. A case study on residential buildings shows the classification accuracy for the existence of doors using ResNet-50 is on average 96.6%, but the interconnectivity of each room has limitations. Pix2Pix-GAN enhances the interconnectivity of each space compared to the door generation results using ResNet-50. Post-processing that combines ResNet-50 and Pix2Pix-GAN has shown an enhanced accuracy of door generation by 16.54% compared to Pix2Pix-GAN alone. In addition to determining door existence, the interconnectivity between all rooms has also been improved by integrating the two models. These results can assist architects in their decision-making process by automatically generating door layout alternatives that take into consideration the spatial interconnectivity.