Numerical simulations of inelastic mechanical behavior of heterogeneous quasibrittle material are essential in a number of engineering areas. Specifically in civil engineering they are applied in assessment of critical buildings and infrastructure made of concrete. Robust high-fidelity models require detailed information about material composition which makes them computationally prohibitive. The projects solves this burden by state of the art techniques using machine learning combined with homogenization and domain decomposition. The heart of the project is advanced physically constrained machine learning surrogate capable to locally replace material response. Such artificial intelligence will be used to emulate behavior of (i) spherical representative volume element in the asymptotic expansion homogenization and (ii) regions tiling the decomposed domain. For such goal the homogenization will be equipped with auxiliary modes accounting for strain localization and the domain decomposition will be improved by reduced order modeling of the communication between individual regions.

Keywords
mechanical behavior;heterogeneity;fracture;strain localization;homogenization;domain decomposition;surrogate;machine learning

Key words in Czech
mechanické chování;heterogenita;lom;lokalizace deformace;homogenizace;doménová dekompozice;náhradní model;strojové učení