| dc.description.abstract | Accurate and rapid seismic response prediction of reinforced concrete (RC) structures in earthquake-prone regions is an important topic in structural and earthquake engineering. However, existing physics-based modeling approaches do not have a good compromise between predictive performance and computational efficiency. High-fidelity models have reasonable predictive performance but are computationally demanding, while more simplified models may be computationally efficient, but do not have as good of performance. The research presented herein aims to address this challenge by developing a novel data-driven computational paradigm via the coupling of machine learning (ML) methods and physics-based models. The ML methods can directly link the experimental data to nonlinear properties of target component, while the physical models meeting universal laws (e.g., Newton’s law of motion) can be used to perform the seismic analysis. Additionally, in real-world scenarios, the dataset is most likely corrupted by outliers, contains missing values, and has sample bias due to the potentially small size. The performance of existing ML methods will be negatively affected by these data-related problems. Thus, novel computational methods to deal with these data-related problems are also developed to make the proposed data-driven framework robust under these circumstances. In sum, the contributions of this dissertation are the following:
1) Two RC column databases, one for rectangular and another for circular columns, were developed.
2) A new ML-based backbone curve model (ML-BCV) was developed by integrating a multi-output least squares support vector machine for regression (MLS-SVMR) with a grid search algorithm for rapid prediction of the bi-linear cyclic backbone curve of RC columns.
3) A novel, locally-weighted ML model (LWLS-SVMR) was developed by combining LS-SVMR and a locally weighted learning algorithm for generalized drift capacity prediction of RC columns.
4) A new, component-level, data-driven framework was developed for generalized, accurate, and efficient seismic response history prediction of structural components subjected to both displacement-controlled cyclic loading and dynamic ground motions. The framework was illustrated for RC columns.
5) The component-level data-driven framework was extended to the system level by coupling it with the simplified, physics-based shear building model. The proposed system-level framework was illustrated for RC frames.
6) A novel, robust, locally-weighted ML model (RLWLS-SVMR) was developed by introducing a weight function into the reformulation of LWLS-SVMR to eliminate the negative effect induced by outliers.
7) A new multiple imputation (MI) method (SRB-PMM) was developed by using sequential regression and predictive mean matching to generate several candidates for imputing (filling in) each missing value while considering the uncertainty associated with the missing data.
8) A novel, regression-based, transfer learning model (DW-SVTR) was developed by coupling two weight functions with LS-SVMR to reduce the negative effect of sample bias due to small datasets. | en |
