Data-Driven Prediction of Out-of-Plane Force Capacity in Unreinforced Masonry Walls Under Blast Loading - A Modeling Approach Driven by Machine Learning

Abstract

Unreinforced masonry walls compromises a large part of the heritage buildings and older civilian buildings in city centers around the world. The current analytical evaluation methods for the out-of-Plane force capacity remain unreliable, which complicates assessments of buildings in regards to updated strength requirements. Masonry walls consist of alternating layers of brick and mortar which makes the dis cretization into a numerical model more complicated. The interfaces and contacts between materials increase rapidly with added layers resulting in very computa tionally heavy models. There are therefore a need for a simplified approach to approximate the force capacity, which may be achieved through the application of machine learning. Numerical modeling within this thesis utilizes a 2D micro-model, where both brick and mortar are included to capture crushing, in any of the materials. The explicit dynamic solver in Abaqus is used to design a quasi-static four point bending test. Geometric and material properties, along with global force and stiffness parameters, are identified and parameterized to generate a comprehensive set of loading cases. A dataset containing the parameters for each loading case of the unreinforced masonry wall is created for machine learning purposes. In this study, several machine learning models are developed with the common objective of predicting the peak force capacity of unreinforced masonry walls. The developed models differ in their prediction targets, including direct prediction of the peak force capacity, prediction of the force–displacement response up until the peak, and the discovery of a new analytical expression for the peak force capacity. The report is written in English.