Architectural space classification considering topological and 3D visual spatial relations using machine learning techniques

Abstract

The paper presents a novel method for classifying architectural spaces in terms of topological and visual relationships required by the functions of the spaces (where spaces such as bedrooms and bathrooms have less visual and physical relationships due to the privacy, while common spaces such as living rooms have higher visual relationship and physical accessibility) through machine learning (ML). The proposed model was applied to single and two-storey residential plans from the leading architects of the 20th century Among the five different ML models whose performances were evaluated comparatively, the best results were obtained with Cascade Forward Neural Networks (CFNN), and the average model success was calculated as 93%. The features affecting the classification models were examined based on SHAP values and revealed that width, control, 3D visibility and 3D natural daylight luminance were among the most influential. The results of five different ML models indicated that the use of topological and 3D visual relationship features in the automated classification of architectural space function can report very high levels of classification accuracy. The findings show that the classification model can be an important part of developing more efficient and adaptive floor plan design, building management and effective reuse strategies.