Seismic attenuation model using artificial neural networks

This research aims at deriving a simple yet powerful ground motion prediction model for the Himalayas and Indo-Gangetic plains, which has a high probability of damage due to moderate or great earthquake, utilizing an entirely data-driven methodology known as an artificial neural network (ANN) from the earthquake recordings of program for excellence in strong motion studies (PESMOS) and central Indo-Gangetic plains network (CIGN) database. Apart from the widely used independent input parameters, magnitude and distance, we have introduced two new variables, focal depth and average seismic shear-wave velocity from the surface to a depth of 30 m (V-s30) to improve model predictability. Peak ground acceleration (PGA) and pseudo-spectral acceleration (PSA) at twenty-five periods are chosen as response variables. The network architecture consisting of 9 hidden nodes were found optimal for the selected database and input-output mapping. The performance of the model is ascertained based on the standard deviation of the error, and the ground motion predictability is tested using real recordings at eight stations and the corresponding widely used ground motion prediction equations (GMPEs) for the Himalayas and Indo-Gangetic Plains. The GMPEs considered for comparison are National Disaster Management Authority (2011), Raghukanth and Kavitha (2014), Anbazhagan et al. (2013), Muthuganeisan and Raghukanth 2016, and Singh et al. (2017). The total standard deviation of response variables in log10 units varies between 0.267-0.343 with period and model predictability plots shows that the current ANN model is competent to predict the response spectrum with good accuracy in both the seismically critical regions of India. Finally, the model is scripted into MATLAB and Excel and supplemented with this article for further use.