Spectral acceleration prediction for strike, dip, and rake: a multi-layered perceptron approach

A multi-layer perceptron (MLP) technique is used to train on the response spectra for various strike angles, dip angles, and rake angles. Fixing the magnitude and depth of the earthquakes, the 3-component ground motion is simulated with the help of SPECFEM3D. The residuals of spectral acceleration as a function of time period, for low-rise to high-rise structures, are found to be free of any trend. The hidden layers in the MLP learn the interdependency of focal mechanism parameters on the response spectrum. The resultant model was checked for attenuation characteristics with respect to distance. Furthermore, the trained MLP also showed a shift in spectral peak due to radiation damping, as expected. This MLP architecture presented in this work can be broadly extended to predict the response spectrum, at bedrock level, for any focal mechanism parameters, i.e., strike, dip, and rake, depending on the velocity model of that region.