Machine learning–based modelling of pseudo spectral acceleration (PSA) using strong-motion data from Türkiye

Abstract In this study, Machine Learning (ML)-based models were developed for the estimation of Pseudo Spectral Acceleration (PSA), a key parameter for identifying strong ground motion in seismic hazard assessments. A large dataset was generated using 1975 earthquake records ( 3.0 ≤ M ≤ 6.8), sourced from 63 strong ground motion stations in different regions of Türkiye. This dataset includes a total of 150,552 samples, which we divided into three groups based on their period (T) values. These 3 datasets were modeled by using 19 different ML algorithms and validated by using tenfold cross-validation. The performances of the models were compared by using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R-squared (R 2 ) criteria. The models trained with the Ensemble Bagged Trees and Fine Tree algorithms were determined to be the best models according to test criteria. The Ensemble Bagged Trees model performed best in the T = 0.05–1.0 s range (R 2 = 0.90), while the Fine Tree model performed best in T = 1.0–1.9 s (R 2 = 0.94) and T = 2.0–4.0 s (R 2 = 0.89). Residual analyses showed that the prediction errors were randomly distributed around the zero line. Moreover, the PSA results of trained ML Models yielded results closer to the observed PSA curve. These results indicate that, within the compiled dataset, the evaluated ML models can predict PSA with lower errors and may serve as a complementary approach to GMPEs in future applications.