Deep generative modelling for nonlinear analysis and in-situ assessment of masonry using multiple mechanical fields

Abstract

Design and in-situ assessment of masonry structures is a challenging task due to the brittle and nonlinear nature of this widely used material, the complex interaction between its components, and the vast variability of material properties in its design space. Current methodologies often rely on oversimplified assumptions that inadequately capture the true mechanical behaviour of masonry or require extensive knowledge and expertise for reliable implementation or interpretation of the obtained results. To overcome these limitations, this article presents an innovative generative machine learning approach, based on conditional generative adversarial network (cGAN), that allows establishing a direct or reverse link between masonry meso-structure and multiple mechanical fields without any specific knowledge of the properties or constitutive laws. The developed cGAN model interprets relationships among multiple mechanical maps using a single model, which leads to enhanced predictions in both linear and nonlinear stages for a wide range of unseen scenarios. The model shows an excellent capability to capture the effect of local and global variability of material properties, constituents sizes, and loading scenarios on the results in both direct (i.e. predicting the strain maps from meso-structure and material properties) and reverse (i.e. predicting the meso-structure and material properties from strain maps) problems. The proposed cGAN modelling approach emerges as a versatile tool with potential broad applications for the design and evaluation of nonlinear composite materials and mechanical behaviour of materials in general, addressing a wide spectrum of engineering challenges.