Design of masonry structures using conditional generative adversarial networks fused with property text information

Abstract

The preliminary design process for masonry structures requires engineers to iteratively verify code compliance and manually develop structural layouts. While image-to-image translation models can automate layout synthesis, existing approaches fall short in incorporating material properties (e.g., compressive strength, rebar yield stress) as explicit design constraints—a critical limitation for structural engineering applications. This study addresses this gap by proposing three fusion architectures that integrate architectural layouts with material property constraints: Direct-GAN (early channel concatenation), Dense Fuse-GAN (bottleneck dense embedding), and Multiscale-GAN (multi-scale skip connection fusion). All models were trained on paired architectural-structural layout datasets and evaluated using perceptual quality metrics (e.g., peak signal-to-noise ratio, structural similarity index measure) and distribution-based measures (e.g., Fréchet inception distance, mean squared error). We report that the Direct-GAN architecture demonstrates superior performance across pixel-level reconstruction accuracy and, hence, can establish an efficient framework for property-aware, data-driven masonry design that advances automation in preliminary structural design workflows.