ALBERT

Description: Our research uses a supervised machine learning algorithm (XGBoost) and strong-motion earthquake recordings from the available seismic networks in India to create a simple but powerful ground motion attenuation model for the seismically hazardous Himalayas, Indo-Gangetic plains, and Kachchh (Gujarat). Our dataset includes 564 Peak ground acceleration (PGA) and pseudo-spectral acceleration (PSA) from 145 events from 75 three-component strong-motion accelerographs in India and surrounding areas. Moment magnitudes, epicentral distances, focal depths, and Vs30 (average seismic shear-wave velocity from the surface to 30m) are our input parameters, and PGA and PSA at twenty-five periods are our output variables. Our XGBoost model predicts ground motion using the entire dataset and a 30% randomised test dataset. Our predictability on the randomised test dataset is 0.986, and the correlation coefficient of observed and predicted PGA values for the whole dataset is 0.998. Our XGBoost ground-motion attenuation model predicts PGA and PSA datasets at four Himalayan and three Kachchh (Gujarat) stations for four events of Mw7.9, Mw6.8, Mw6.2, and Mw5.6, suggesting it would work for the Himalayas, Indo-Gangetic plains, and Kachchh (Gujarat). Another GMPE model was created using the LightGBM gradient boosting algorithm on our dataset. Finally, we compared our XGBoost model predictions to the three Himalayan GMPEs and LightGBM. The XGBoost model predicted ground motion most accurately. Period-dependent log10 output variable standard deviations range from 0.459 to 0.780. Our model predictability plots show that the current ML model can accurately predict the response spectrum for seismically high-hazard zones of India.