Peak ground acceleration prediction using supervised machine learning algorithm for the seismically hazardous Kachchh rift zone, Gujarat, India

The lack of essential input parameters and the utilisation of the least squares technique have led to a significant amount of uncertainty in the predictive model of empirical ground motion prediction models (GMPEs) in India. These models are primarily based on epicentral distances and magnitudes. In the present study, we employ a supervised Machine Learning approach known as XGBoost to develop a reliable and robust GMPE for estimating peak ground acceleration (PGA) in Kachchh, Gujarat. Our methodology utilises various independent input parameters, including moment magnitudes, hypocentral distances, focal depths, and site proxy represented by the average seismic shear-wave velocity from the surface to a depth of 30 m (V s30). The research utilised a dataset of seismic records, specifically eight engineering seismoscope (SRR) records from the 2001 M w7.7 Bhuj mainshock and 237 strong-motion records from 32 notable Bhuj aftershocks of M w3.3–5.6 (occurring between 2002 and 2008). These records were collected at various epicentral distances, ranging from 1.0 to 288 km. The ground motion predictability of our XGBoost model is assessed by conducting tests on both the entire dataset and a randomised test dataset comprising 30% of the total data points. The correlation coefficient between the observed and predicted PGA values for the entire dataset is determined to be ± 0.994. Conversely, when examining the predictability of the XGBoost model on the randomised test dataset, we observe a correlation coefficient of 0.844. The XGBoost model demonstrates superior predictive performance compared to the previously developed GMPEs for Kachchh, Gujarat. These GMPEs were derived using artificial neural networks and nonlinear least squares methods, utilising the same PGA dataset. Additionally, the developed XGBoost model exhibits a high level of predictability, indicating its ability to accurately forecast the observed PGA dataset for four earthquakes with magnitudes ranging from M w5.6 to M w7.7. These earthquakes include the 2001 M w7.7 Bhuj (Gujarat) mainshock, the 1999 M w6.6 Chamoli mainshock, and two M w5.6 Bhuj aftershocks in 2006. This suggests that the XGBoost ML GMPE model developed in this study holds potential for effectively estimating ground motions for earthquakes occurring in Kachchh, Gujarat, India. Furthermore, it may also prove valuable in predicting ground motions for earthquakes in Uttarakhand, India. The findings of our study indicate that the classification of sites into different classes based on varying soil types, with a V s30 range of 180–1500 m/s, has a substantial impact on the attenuation curves predicted by the XGBoost model. The XGBoost model is also implemented in Python and included in this article for future utilisation.