PREDICTION OF PEAK GROUND ACCELERATION OF SHALLOW CRUSTAL EARTHQUAKES USING DEEP NEURON NETWORK

Ground motion prediction equations (GMPEs) can provide estimates of peak ground motion parameters (e.g., PGA,) by considering the effects of earthquake magnitude, distance, site, and source, which is a crucial element in the seismic hazard analysis. These equations help to evaluate the mean/median ground shaking effects for the expected levels of earthquakes. Thus, many empirical GMPEs were developed based on regression analysis, but they are highly uncertain due to the uncertainties of independent variables. To overcome this issue, the artificial neural network (ANN) is applied in the field of seismology and earthquake engineering to predict ground motion intensity measures (IMs). However, even though the ANN can produce high prediction accuracy, the defects are apparent, such as the database is relatively small. As a result, a challenging problem that arises in this domain is whether an artificial intelligence-based method can predict the attenuation relationship of ground motion parameters as well as response spectra based on a big database, which has not been reported elsewhere. Recently, deep neural network (DNN) has received much attention from engineering and is playing a crucial role in providing big data predictive models. Therefore, this study develops a DNN trained by the recordings from the PEER NGA-West2 database to predict peak ground motion acceleration (PGA). To this end, we collect 20,900 GMs based on the proposed criteria from the PEER NGA-West2 database and randomly split them into the training, validation, and testing datasets. The developed model relates PGA to earthquake source to site distance, earthquake magnitude, average shear-wave velocity, faulting mechanisms, and focal depth. The prediction errors are evaluated by three performance indicators, and the predictive results are compared with five well-known empirical models and one artificial neuron network model developed based on the PEER NGA-West2 database. The between-event and within-event residuals are calculated and compared with other models. Based on the results, our model has the best goodness-of-fit statistics of all the GMPEs we have compared, confirming that the proposed model is associated with better predictive power.