Date of Award

Summer 5-31-2024

Access Type

Thesis - Open Access

Degree Name

Master of Science in Civil Engineering

Department

Civil Engineering

Committee Chair

Siddharth Parida

First Committee Member

Prashant Shekhar

Second Committee Member

Jeff R. Brown

Third Committee Member

Dan Su

Abstract

The high computational cost of estimating engineering demand parameters (EDPs) via finite element (FE) models, which incorporate uncertainties in earthquake events and material properties, limits the application of the Performance-Based Earthquake Engineering (PBEE) framework. Previous efforts to replace FE models with surrogate models have typically focused only on building parameters, necessitating re-training for new, unseen earthquakes. This paper introduces a machine learning-based surrogate model framework that addresses both earthquake and material parameter uncertainties to predict responses for unseen seismic events. Earthquakes are characterized by their projections on an orthonormal basis, computed using Singular Value Decomposition (SVD) of a representative ground motion suite, allowing for the generation of varied earthquake scenarios by sampling these weights. These weights, along with constitutive parameters, serve as inputs to the machine learning models, with EDPs as the output. Four competing machine learning models were evaluated, with deep neural networks (DNNs) demonstrating the highest accuracy. The framework's validity is shown through its successful prediction of the peak responses of one-story and three-story shear frame buildings, represented as nonlinear spring–mass–damper systems, subjected to unseen far-field ground motions. Furthermore, the study highlights the importance of rigorously characterizing forcing functions, as predictions are highly sensitive to these parameters. Machine learning tools, due to their flexibility and efficiency, have emerged as powerful alternatives in various engineering fields. In this study, the application of machine learning for predicting EDPs, while considering uncertainties in both forcing functions and model parameters, is assessed. Results indicate that DNNs perform the best among the tested models. This comprehensive framework integrates machine learning into the PBEE framework, offering a cost-effective solution for structural analysis under uncertain conditions.

Share

COinS