Intelligent generation and interpretability analysis of shear wall structure design by learning from multidimensional to high-dimensional features

Abstract

The intelligent design of shear wall structures is a critical aspect of smart construction, with a high demand for research and applications. Accurately predicting the shear wall ratio (i.e., the shear wall area-to-floor area ratio) during cost estimation and rapidly generating shear wall layouts during early design is essential. However, the unclear influences of numerous design feature parameters hinder the enhancement of generative AI design. This affects both the prediction of shear wall ratios from multidimensional features and the generation of shear wall layouts from high-dimensional features. Therefore, a method for generating key structural design features using machine learning (ML) and generative adversarial networks (GANs), along with model interpretation, is proposed in this study. Existing shear wall design data are collected, and features such as the architectural plan geometry, seismic design conditions, and shear wall ratios are extracted to establish a dataset. Key shear wall ratio parameters are predicted using an ML model with multidimensional design features as inputs, and interpretability analysis is conducted using Shapley Additive Explanations (SHAP). Concurrently, a GAN model is built to generate shear wall designs using fused image-text high-dimensional features, and the influence patterns of design features are explained through sensitivity analysis. The analysis results indicate that the prediction accuracy is effectively enhanced by ML-based multidimensional feature learning, shear wall designs are effectively generated by GAN-based high-dimensional feature learning, and seismic design intensity and structural height are revealed as significant factors through interpretability analysis. Furthermore, when high-dimensional feature inputs are available, the generation of comprehensive features should be prioritized for shear wall structural designs.