Quantifying Vernacular Window Geometry via CNNs for Simulation-Oriented Analysis: An Iranian Architecture Case Study

Environmental performance is a key concern in contemporary architectural design, as buildings significantly affect energy consumption and contribute to climate change. Iranian vernacular architecture, particularly traditional window types, incorporates passive design strategies associated with thermal comfort regulation and daylight modulation. Decoding these window typologies can reveal climate-responsive design principles rooted in environmental and cultural contexts, with continued relevance to current sustainability goals. However, the geometric and formal complexity of traditional Iranian window systems has limited the systematic analysis of these systems using conventional methods. Existing research has primarily relied on visual, descriptive, or syntactic approaches, with limited emphasis on extracting quantitative, simulation-oriented data. This study proposes a convolutional neural network (CNN)-based visual framework for automated classification of traditional window typologies and extraction of architectural parameters from façade images. Fine-tuned CNNs are employed to identify window types and derive key geometric descriptors commonly used in environmental performance modeling, including window-to-wall ratio, frame depth, and perforation density. The extracted parameters are organized into structured, simulation-oriented data sets intended for downstream use in energy and daylight modeling workflows, including tools such as EnergyPlus and Radiance. Gradient-weighted class activation mapping is used to verify that the model focuses on morphologically meaningful features, while statistical analysis (Kruskal–Wallis, P < 0.001) confirms significant typological differentiation in the extracted parameters. While environmental simulations are not conducted in the present study, the resulting parameter sets are explicitly designed to be compatible with established simulation platforms, enabling downstream use in performance evaluation workflows. Although developed using traditional Iranian window typologies, the proposed framework provides a transferable methodology for linking visual data with simulation-oriented architectural parameterization.