Earthquake response of structures via reduced-order surrogates

Almasarani, Omar

dc.contributor.author	Almasarani, Omar	en
dc.date.accessioned	2022-03-08T09:48:53Z
dc.date.available	2022-03-08T09:48:53Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/54938
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.22636
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	Surrogate modeling	en
dc.subject	Proper Orthogonal Decomposition (POD)	en
dc.subject	Structural dynamic analysis	en
dc.subject	Machine learning	en
dc.subject	Earthquake engineering	en
dc.title	Earthquake response of structures via reduced-order surrogates	en
heal.type	masterThesis
heal.classification	Structural engineering	en
heal.language	en
heal.access	campus
heal.recordProvider	ntua	el
heal.publicationDate	2021-10-29
heal.abstract	This thesis focuses on reducing the computational cost of high-fidelity structural analyses using numerical methods such as finite element method (FEM) by using surrogate models techniques of 3D structures for earthquake analysis to alleviate such computational burden of estimating the seismic response for structures. This done by addressing two surrogate models, this first one is analytical approach using Model Order Reduction (MOR), i.e approximate the original full high-order problems to a system of lower dimension. The second approach is using Machine Learning techniques and use the trained models to predict the response the dynamic response of the structure, i.e., the maximum displacement. The first part of this thesis studies the application of Proper Orthogonal Decomposition (POD) to reduce the full-order of linear dynamic models of 2D and 3D frames under seismic excitation and investigate its efficiency in constructing Reduced Order Models (ROM) of the dynamic response of the structures. This is done by, first apply Newmark's method to estimate the dynamic response the full-order model and pick a small discrete time (snapshots) from the solution. Then, truncate perform a Singular Value Decomposition to determine a low-dimensional approximation to high-dimensional model in terms of dominant patterns then by getting a reduced rank of equation of motion (EOM), perform the Newmark's analysis. Finally Having the reduced solutions, map back the the full rank system. The performance of POD is assessed by comparing the results of the reduced systems to the original ones using a using MATLAB script that was developed to perform all the analyses of linear systems and the procedure of POD method. The second part of this thesis is Machine Learning techniques i.e., Shallow/Deep Neural Network and Support Vector Machines, each algorithm was used to get trained models to predict the dynamic response of the structure. a large dataset of a ground motions (GM) was collected according to the most common recorded earthquakes in the literature and select the Intensity Measures of the GMs i.e, Peak Ground Acceleration and Spectral Acceleration as input features training dataset along with and the predominant period of the structures, to predict the maximum displacement of two separate cases that were addressed. The first case, to predict the response of single structures, and the second case, is to predict the response a group of structures together. The labels of the training dataset were obtained from the dynamic analysis of linear system using Newmarks's method. The selection of the optimum algorithm was based on the error values obtained from training, validation and testing. Also, a comparison between the methods efficiency according to the prediction accuracy and training time was investigated. Finally, addressing the importance of selection the appropriate input features to get the best accurate outcome. Sap2000 was used to creates all the FE 3D models that were investigated in this project, and then the mass and stiffness matrices were extracted to a MATLAB scripts for each case. For the POD case, a MATLAB script was developed according to the literature. And for Machine Learning cases, the MATLAB scripts were generated from the built-in toolboxes with modifying parameters to suit each case accordingly.	en
heal.advisorName	Triantafyllou, Savvas	en
heal.committeeMemberName	Papadopoulos, Vissarion	en
heal.committeeMemberName	Vamvatsikos, Dimitrios	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Πολιτικών Μηχανικών	el
heal.academicPublisherID	ntua
heal.numberOfPages	95 σ.	el
heal.fullTextAvailability	false