A machine learning based material homogenization technique for masonry structures

Cutting-edge methods in the computational analysis of structures have been developed over the last decades. Such modern tools are helpful to assess the safety of existing buildings. Two main finite element (FE) modeling approaches have been developed in the field of masonry structures, i.e. micro an...

Descripción completa

Detalles Bibliográficos
Autor: Kalkbrenner, Philip
Tipo de recurso: tesis doctoral
Fecha de publicación:2021
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/360902
Acceso en línea:https://hdl.handle.net/2117/360902
https://dx.doi.org/10.5821/dissertation-2117-360902
Access Level:acceso abierto
Palabra clave:Masonry
Finite element method
Macro scale modeling
Micro scale modeling
Multi scale modeling
Material homogenization
Heterogeneous material
Machine learning
Continuum damage mechanics
Pushover analysis
Palacio Pereira
Àrees temàtiques de la UPC::Enginyeria civil
Descripción
Sumario:Cutting-edge methods in the computational analysis of structures have been developed over the last decades. Such modern tools are helpful to assess the safety of existing buildings. Two main finite element (FE) modeling approaches have been developed in the field of masonry structures, i.e. micro and macro scale. While the micro modeling distinguishes between the masonry components in order to accurately represent the typical masonry damage mechanisms in the material constituents, macro modeling considers a single continuum material with smeared properties so that large scale masonry models can be analyzed. Both techniques have demonstrated their advantages in different structural applications. However, each approach comes along with some possible disadvantages. For example, the use of micro modeling is limited to small scale structures, since the computational effort becomes too expensive for large scale applications, while macro modeling cannot take into account precisely the complex interaction among masonry components (brick units and mortar joints). Multi scale techniques have been proposed to combine the accuracy of micro modeling and the computational efficiency of macro modeling. Such procedures consider linked FE analyses at both scales, and are based on the concept of a representative volume element (RVE). The analysis of a RVE takes into account the micro structural behavior of component materials, and scales it up to the macro level. In spite of being a very accurate tool for the analysis of masonry structures, multi scale techniques still exhibit high computational cost while connecting the FE analyses at the two scales. Machine learning (ML) tools have been utilized successfully to train specific models by feeding big source data from different fields, e.g. autonomous driving, face recognition, etc. This thesis proposes the use of ML to develop a novel homogenization strategy for the in-plane analysis of masonry structures, where a continuous nonlinear material law is calibrated by considering relevant data derived from micro scale analysis. The proposed method is based on a ML tool that links the macro and micro scales of the analysis, by training a macro model smeared damage constitutive law through benchmark data from numerical tests derived from RVE micro models. In this context, numerical nonlinear tests on masonry micro models executed in a virtual laboratory provide the benchmark data for feeding the ML training procedure. The adopted ML technique allows the accurate and efficient simulation of the anisotropic behavior of masonry material by means of a tensor mapping procedure. The final stage of this novel homogenization method is the definition of a calibrated continuum constitutive model for the structural application to the masonry macro scale. The developed technique is applied to the in-plane homogenization of a Flemish bond masonry wall. Evaluation examples based on the simulation of physical laboratory tests show the accuracy of the method when compared with sophisticated micro modeling of the entire structure. Finally, an application example of the novel homogenization technique is given for the pushover analysis of a masonry heritage structure.