Data-driven ground motion model for aftershock seismic hazard assessment in the Himalayan region

Aftershock Earthquake poses a significant risk to structures already weakened by a mainshock, potentially leading to further damage or collapse. Despite their critical impact, aftershocks are often overlooked in seismic design practices due to the limited availability of comprehensive mainshock-aftershock datasets. Existing ground motion models (GMMs) primarily focus on mainshock events and do not fully capture the distinct characteristics of aftershock sequences. To address this gap, this study aims to develop a GMM for 5 ​% damped elastic aftershock response spectra for the Himalayan region using an artificial neural network (ANN) algorithm. The dataset contains numerous sequences including the 1999 Chamoli, 2011 Sikkim, 2013 Uttarkashi, and 2015 Nepal earthquakes, supplemented by records from the NGA-West2 database. Mainshock spectral ordinate and other parameters such as the difference between mainshock and aftershocks magnitude, difference between their hypocentral depth, logarithm of mainshock to aftershocks representative distance ratio, ratio of mainshock to aftershocks magnitude ratio to logarithm of mainshock rupture distance and a site classification flag are taken as predictors. Distinct effect of all the considered predictor variables on the aftershock ordinates are observed, with the most important predictor being the mainshock Spectral ordinates followed by difference between hypocentral depth of mainshock and aftershock. The model achieved high accuracy in predicting aftershock spectra, with a coefficient of determination (R 2 ) of 0.861, mean square error (MSE) of 0.078 and mean absolute error (MAE) of 0.215. The model outperforms existing global GMMs in capturing aftershock spectral characteristics, particularly in the Himalayan region. Mixed-effects regression framework is adopted to account for inter- and intra-event variabilities and the respective standard deviations are reported. Residual analysis confirms the robustness of model and demonstrated minimal bias. The total standard deviation is reported to lie between 0.268 and 0.488.