PREDICTION OF STRONG GROUND MOTION PARAMETERS USING MACHINE LEARNING TECHNIQUES

Fahrettin Kuran

(Thesis Supervisors: Assoc. Prof. Gülüm Tanırcan and Assoc. Prof. Elham Pashaei )

ABSTRACT

Peak ground velocity (PGV) and cumulative absolute velocity (CAV) are powerful intensity measures for quantifying potential earthquake damage. Reliable prediction of those parameters is of essential importance in the precise calculation of seismic hazard. Machine learning can provide accurate and reliable predictions of PGV and CAV due to handling nonlinear relationships, adaptability to changing conditions, automation, efficiency, and the potential for real-time predictions. This study aims to comprehensively compare the relative performance of six different machine learning algorithms for PGV and CAV prediction and develop the Turkiye-specific ground motion models with the most recent strong motion dataset. Support Vector Machine, Linear Regression, Random Forest, Artificial Neural Network, Bayesian Ridge Regression, and Gradient Boosting algorithms are evaluated and compared. The Turkish strong motion database, which consists of over 950 earthquakes occurring from 1983 to 2023, and a worldwide near-field earthquake database (moment magnitude≥5.5) are used for shaping the models' ability to learn and make accurate predictions. Various source, site, distance, and faulting parameters are considered as estimator parameters. Three feature selection methods, embedded, filter, and wrapper, are applied to determine the most suitable estimator parameters to predict PGV and CAV. Statistical evaluation metrics are employed to measure the performance of the models. Among algorithms, the Gradient Boosting algorithm shows remarkable success in the prediction of both PGV and CAV. PGV prediction performance is better when all estimator parameters are used together, while CAV prediction ability is better when user-selected parameters are utilized. Outlier elimination processes are found to be redundant in predicting the PGVs and CAVs of large earthquakes. The proposed PGV and CAV models are applicable to shallow crustal strike-slip and normal faulting earthquakes with moment magnitude ranging from 3.5 to 7.8 and Joyner and Boore distance up to 200 km.