A neural network-based ground motion model trained on ground motion simulations in New Zealand

Abstract

This poster presents a New Zealand-specific simulation-based ground motion model (GMM) developed using an artificial neural network (ANN). The rationale behind the model development is to leverage the advantages of physics-based ground motion simulations in a manner that is computationally tractable for use in probabilistic seismic hazard analysis (PSHA), when a large number of seismic sources are considered, principally in the representation of background, area, and distributed seismicity sources; as well as another means for synthesis of simulation results. We demonstrate the development of such a model using simulation data from both validation of historical events (Lee et al. (2020); approximately 1000 simulations), as well as future potential events ((Motha et al. (2020); approximately 17,000 simulations) in New Zealand. A multi-layer perceptron ANN with multiple hidden layers was used and trained using the back-propagation algorithm. Evaluation was done on 10% of the simulated events, selected randomly. To illustrate the predictive capabilities of the model we compare the results obtained against empirical GMMs, as well as examining the predicted variation in ground motion amplitudes as a function of features such as source and site location, tectonic type etc. The principal caveat of this approach is that it is limited by the validity of the underlying simulations, however, as these continue to improve the developed models can be iteratively improved to provide a seamless means for use in simulation-based PSHA.