Numerical analysis of large masonry structures: bridging meso and macro scales via artificial neural networks

Abstract

This paper presents a methodology for analysis of large-scale masonry structures. The approach involves development of a series of artificial neural networks which enable the identification of main variables employed in the macroscopic formulation that incorporates an inelastic constitutive law with embedded discontinuity. The data required for training of neural networks is generated using ‘virtual experiments’, whereby the ‘equivalent’ anisotropic response of masonry is obtained through a mesoscale finite element analysis of masonry wallets. The paper outlines the procedure for identification of approximation coefficients describing the orientation-dependency of strength, and other relevant parameters. A numerical example is provided involving analysis of a large masonry wall with multiple openings. The results of macroscale approach are compared with those based on a detailed mesoscale model for the same geometry and boundary conditions.