Architectural section generation and semantic evaluation with deep learning methods

Abstract

Architectural section drawings serve as crucial visual tools in the fields of architectural education and professional practice, allowing for a deeper understanding and communication of spatial configurations. Within this classification, orthographic sections stand out for their ability to depict various levels within a structure, showcase the intricate relationships between different spaces, and detail the essential structural elements. Section drawings provide a glimpse into the design, enabling architects to understand how various architectural elements interact and coexist within the overall architectural composition. Throughout architectural history, plan drawings have dominated representation, compared to section drawings, which often require interpretive processes such as documentation (Guillerme and Verin, 1989). Although the definitions and roles of section drawings have evolved throughout architectural history, plan drawings have gradually assumed a more dominant position. The prominence of plans is not only well-established in architectural discourse but also widely reflected in computational design practices. In particular, rule-based design approaches tend to foreground plan drawings, often relegating sections to a secondary role. For a focused discussion on the architectural section, this study proposes a case analysis based on the works of Andrea Palladio, the architect whose design principles and built forms have been extensively examined as one of the canons of architectural history. Thesis aims to investigate the potential of machine learning–based methods, particularly deep learning approaches such as semantic segmentation and generative adversarial networks (GANs), in analyzing and interpreting Palladio's architectural logic through the section plane. Palladio's villas, known for their geometric clarity and dimensional consistency, provide a reliable dataset for this investigation. In the initial phase of the thesis, the spatial ratios, height relationships, and organizational features of Palladian architecture were examined through plan and section drawings. While plan drawings often conform to systematic geometric rules, section drawings present more complex challenges, as they incorporate not only walls, openings, and columns but also ceilings, roofs, and vertical spatial information. At the first stage, Palladio's section drawings were examined, and the results revealed that Palladian sectional drawings do not follow the specific rules that plan drawings do. Space organization can vary depending on the sectioned point. While architectural elements reflect this change, it is possible to distinguish sectional drawings from plan drawings. Palladio's plan drawings, as Stiny and Mitchell (1978) point out, may consist of specific sets of rules; sectional drawings, on the other hand, are not based on these rules, but rather are created by replicating architectural elements and components according to specific principles. While section drawings offer different possibilities, doing a rule-based study is limited. Therefore, in the continuation of the study, various generative networks were employed to generate designs that are not based on rules but exhibit repetitive similarities. Semantic segmentation was applied to understand the section drawings better and to distinguish between plans. Semantic segmentation is a model that assigns each pixel in an image to a meaningful class. Classes of architectural elements were created using semantic segmentation using the U-Net model. This approach aimed not to create new drawings, but to reveal underlying spatial patterns by analyzing and analyzing existing sectional data. Despite the limited dataset, this method successfully distinguished between different spatial and structural components within section drawings, yielding diverse results. This process was similarly applied to plan drawings. Comparing the results of the two different datasets revealed the class diversity in the section dataset and, consequently, the higher accuracy of predictions. The thesis also aimed to test various generative models to facilitate image-based generation. For this phase, a dataset with high-dimensional data was required. After demonstrating the difference between a section drawing and a plan, the goal was to create a dataset and select the most appropriate model to generate Palladian-style sections. One of the biggest limitations of the thesis was creating a dataset for the section. During the dataset research for semantic segmentation, we discovered that the number of section drawings was small. Even Palladio, in his book on his principles, did not include a section on the entire building. In his book, Palladio often refers to sections as detailed drawings. In this context, a real-world cross-section dataset, specifically derived from Bertotti's drawings, was converted into a 3D model format, taking into account Palladio's plan drawings and principles. Cross-section slices from these 3D models then formed the dataset and were used in the relevant training processes. At the next stage, a comprehensive literature review on architecture related and machine learning was conducted to explore these issues, followed by the application of various models, including StyleGAN, DCGAN, and VAE, to a set of collected real datasets and created datasets. Although these generative models yielded limited accuracy and spatial comprehension results, the thesis found that image-based generative model , such as DCGAN, offered more promising outcomes. While not delivering fully defined sections, these models provided approximations of missing spaces and elements, allowing for architectural reasoning about typologies and spatial components. In doing so, they served as image-generation tools and platforms for architectural speculation and discussion. Considering the analyses of previous models and the nature of the obtained outputs, the DCGAN architecture has been chosen for this phase of the thesis to enhance the level of detail. This study continued the analysis of loss graphs, FID scores, and overall visual outputs using the same methods used in previous phases. Furthermore, to provide a more in-depth analysis of the architectural outputs, these outputs were evaluated based on sagittal, frontal, composite, and color composite datasets. Although all the datasets have issues such as overfitting or mode collapse, certain architectural features from the datasets make evaluation possible. The especially colored composite section dataset, which includes a semantic version of the sections, demonstrates more opportunities for future projects. The overall results of the study demonstrate that examining section drawings based on the part-whole relationship of architectural elements offers a more suitable tool for design processes. Plan-based visual generation more clearly presents organizational structures, while section drawings can provide a more layered and semantically rich representation that encompasses structural narratives. Despite the limited dataset, it has been observed that different deep learning models have generated valuable results about section drawings and have revealed the features that distinguish section drawings from plan drawings. Semantic segmentation, in particular, has emerged as an effective tool for extracting meaningful information from complex visual inputs. Semantic segmentation, which highlights the differences between plan and section through classes, enables us to understand the potential of section drawings better. Experiments conducted on DCGAN demonstrate that, despite generating images with less data, the results obtained with classed colored combined section dataset demonstrate that section studies provide significant information about architectural elements and the part-whole relationship. This demonstrates that section drawings extend beyond mere visual representation and serve as a valuable tool for examining the relationships between structural elements in greater depth. Future work plans to focus more on section-based analysis and develop a new dataset specifically targeting architectural elements. This will enable the detailed analysis of section drawings using deep learning methods to gain a better understanding of the internal structural logic of individual components in architectural systems. Predictive modeling techniques have demonstrated utility in tasks such as completing architectural documentation, whereas image generation models offer promising opportunities for creating innovative design proposals. Consequently, a shift in focus toward data-driven approaches, such as semantic segmentation, which can reveal deeper structural and semantic layers, proved necessary. Such methods enable more reliable analysis of architectural section drawings, contributing to design and digital documentation practices. In conclusion, this research offers a methodological framework for reading, analyzing, and reconstructing architectural sections through machine learning techniques. It contributes to understanding spatial organization and design logic while proposing a new approach for digital documentation of cultural heritage. Additionally, it opens up avenues for further discussions on the use of predictive models in design speculation and restitution processes, particularly in relation to Palladian villas. Future work will focus on developing datasets centered around architectural elements and the part–whole relationship, suggesting new pathways for section-based architectural analysis and synthetic reconstruction.