Machine Learning for Extracting Spatial Logic in Student Design Proposals: A Pattern Language–Based Approach

Machine Learning for Extracting Spatial Logic in Student Design Proposals: A Pattern Language–Based Approach

Gholamhosein Naseri1 Hossein Moradi Nasab2 Kamran Javaheri3 Hana Avani4

1) Department of architecture, Damghan Branch, Islamic Azad University, Damghan, Iran,
2) Department of architecture, Se.C, Islamic Azad University, Semnan, Iran,
3) Department of architecture, Se.C, Islamic Azad University, Semnan, Iran,
4) Department of architecture, Se.C, Islamic Azad University, Semnan, Iran,

IOI Code : AAHD-ABHFCG
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Extracting the latent spatial logic embedded within a vast corpus of student architectural design projects has consistently been a fundamental challenge in architectural theory and pedagogy. Christopher Alexander s Pattern Language theory, despite its profound influence on the field of design, has rarely been formulated within contemporary computational frameworks to enable the automated extraction of patterns from large-scale design data. This research aims to bridge this gap by presenting a machine learning-based method for extracting recurring spatial patterns from student projects. The study adopts a sequential exploratory mixed-methods approach, utilizing Graph Autoencoders alongside Hierarchical and K-Means clustering techniques to map the spatial graphs of 255 architectural design projects from Iranian architecture schools into a latent space, thereby identifying dominant and exceptional pattern clusters. The validation of results has been conducted through the Delphi method and comparative analysis of the clusters against the evaluations of experienced studio instructors. The findings reveal three principal pattern families—namely, Core-Centered, Linear-Hierarchical, and Cluster-Networked —which, while overlapping with Alexander s classical patterns, also exhibit significant differences that reflect the natural evolution of design language within the context of contemporary education. Furthermore, a meaningful convergence was observed between the extracted patterns and the instructors assessment scores, indicating the efficacy of this method in evidence-based evaluation. Additionally, the identification of exceptional and boundary clusters demonstrated that architectural creativity emerges not through a departure from patterns, but through their creative recombination. By establishing a connection between Pattern Language theory and modern computational methods, this research takes a step toward objectifying evaluation and feedback processes in architectural education, offering an efficient tool for instructors, students, and researchers in this field.

Keywords : machine learning pattern language spatial logic architectural education graph autoencoder student design analysis