Using an artificial neural network approach to develop a nonergodic ground motion model: Incorporating systematic spatial variations for Southern California

Abstract:

Path and site effects make up a large proportion of uncertainty in ground motion modeling. Characterizing spatial variations allows the ergodic assumption on ground motion records to be relaxed so that models can become specific to regional event, path, and site terms. Systematic path effects have been constrained in ground motion modeling with varying coefficient models (Landwehr et al., 2016), mixed-effects path term residuals (Sahakian et al., 2019), and gridded anelastic attenuation terms (Dawood & Rodriguez-Marek, 2013; Kuehn et al., 2019). Artificial intelligence techniques have recently increased in number and begun to advance many ground motion modeling applications, particularly in cases with large datasets. Machine learning methods provide an alternative approach to including systematic spatial effects within ground motion models, that can simplify the development of complicated interconnected relationships of parameters.

We create an artificial neural network (ANN) ground motion model including nonergodic spatial variation of coefficients. We expand off the development of the artificial neural network of Withers et al. (2020) to include spatially varying path terms by including additional training predictor variables. We explore various methods of imposing prior constraints on the neural network predictions, including imposing gaussian process layers with spatially smoothed kernels. Additionally, we also experiment with a discrete gridded approach by training on a model with varying sized spatial cells that correspond to the effects of anelastic attenuation. We use a combined dataset of synthetic CyberShake and empirical data from Southern California to train and evaluate our results. Including simulated earthquakes allows for dense spatial coverage of event and station locations, while the recorded database expands the applicable range of our model to a larger range of magnitude and site conditions for validation. The database targets are ground motion residuals with reference to empirically based ergodic ground motion models. This method provides a straight-forward approach to compute nonergodic adjustments to preexisting ground motions models. Finally, our approach allows the estimation of epistemic and aleatory uncertainty, after accounting for adjustments in the median.