We have investigated the artificial neural network method for the derivation of physically sound, easy‐to‐handle, predictive ground‐motion models. Avoiding the specification of any a priori functional form, artificial neural networks (ANNs) provide fully data‐driven predictive models and allow the testing of the relative importance of the effects of independent variables on seismic ground motion. This approach is applied here as an illustrative example, using a large subset of the KiK‐net seismic database, which includes 3891 records from 398 sites and 335 earthquakes. The independent variables tested are the moment magnitude (Mw), the focal depth, the epicentral distance (R), the site resonance frequency (f0), and the time‐averaged shear‐wave velocity down to 30 m (VS30). The neural model output is the horizontal peak ground acceleration (PGA). The standard deviation obtained for the model of 0.34 is comparable to, or slightly lower than, conventional ground‐motion prediction equations (GMPEs). Although not imposed a priori, these results have a number of physically sound features: clear magnitude and depth dependency of the decay of the ground motion with distance, near‐fault saturation for large magnitudes, and indications of nonlinear effects in softer soils. In addition, the ANN method also allows the ranking of the importance of explanatory input parameters: while Mw and R represent the key control parameters, depth is shown to mainly affect moderate magnitude events in the epicentral area. Considering the two site parameters VS30 and f0, the latter is shown to be more efficient in fitting the data than VS30. The ability to implement this model using Microsoft Excel or another simple script is demonstrated, which opens a vast field for its use.
Online Material: MATLAB script and parameters used for ANN model implementation.