Structural Aesthetics
Str
Arch
Comp
A Computational design framework evaluates architecture through two lenses: convergent structural performance and less-explored aesthetic qualities. It provides a diverse array of high-performing, aesthetically tailored solutions. Using Finite Element Analysis (FEA) and Machine Learning (ML), the framework assesses both structural integrity and aesthetic qualities. The process calibrates ML parameters to recognize embedded aesthetics through curated historical architectural data and adapts FEA for cross-sectional image analysis. Focusing on domes in mosques and churches, this study opens a multidimensional design space that integrates structural performance with cultural specificity, marking an initial effort to bring aesthetics into computational workflow
To train a model for recognizing dome cross-sections, historical cross-section images were collected and curated to focus on poché (solid black fill). This curation was automated using a Pix2Pix model, a conditional Generative Adversarial Network (cGAN) that converts historical images into poché-style cross-sections. The curated images were then used to train a Convolutional Neural Network (CNN) classifier for accurate dome cross-section recognition and classification.
Historical Cross-sections Drawings
Input
Ground Truth
Prediction
conditional Generative Adversarial Network (cGAN)
This investigation explored architectural patterns and ornamentation by training a Convolutional Neural Network (CNN) on pattern classification. A diffusion model then generated new patterns, using a specific loss function to differentiate between two pattern classes. The diffusion model adds noise to data and reverses it to produce new samples with targeted stylistic features.
Gothic Patterns (Class a)
Islamic Patterns (Class b)
Historical Ornament Images
+1 Class a | -1 Class b
-1 Class a | +1 Class b
Diffusion Model (Pattern Generation)
A parametric model was developed to generate a diverse dataset of dome shapes, including variations beyond conventional dome cross-sections.
Parametric Model (Dome Cross-section Generation)
These cross-sections were then processed through a custom finite element analysis (FEA) library that operates on a pixel level, converting each pixel into a finite element for analysis. This pixel-based FEA approach enabled efficient batch processing and analysis of images in large numbers. then using the maximum displacement of each iteration to classify it's structural score.
Pixel FEA
Based on the cross-section classifications, the activation of each class was used to shape the loss function for the diffusion model, guiding it to generate patterns that align with each cross-section type. These generated patterns were then mapped onto the domes.
Map Textures
Using these classifications, a multidimensional space of design solutions could be explored.
Latent Space of Design Iterations
