Using Machine Learning to Improve Ground Motion Prediction Equations

Date: 4/24/2019

Time: 08:45 AM

Room: Elliott Bay

I assess the potential for artificial neural networks to capture complex relationships among input parameters in ground-motion prediction equations (GMPEs) that may be difficult to describe in traditional empirical regressions. I construct a suite of GMPEs for California using a convolution artificial neural network model with data from the Pacific Earthquake Engineering Research Center’s Next Generation Attenuation West2 ground-motion database. I consider different parameterizations for capturing the spatial variations in ground motion associated with the earthquake mechanism, site characterization, and distance from the rupture. For the inputs associated with each effect, I assess how the neural network weights the combination of parameters, the effect on the uncertainty in the resulting GMPE, and compare the functional form to traditional empirical regression-based GMPEs. For example, using continuous independent variables, such as fault dip and slip rake, rather than classifying earthquakes into discrete mechanism types (e.g., strike-slip, reverse, and normal faulting) results in slightly smaller residuals. Additionally, complex variations in amplitude associated with hanging wall/footwall effects can be captured using a combination of Joyner-Boore distance and rupture distance. The artificial neural network provides a generalized form that remains the same for all of these different parameterizations, thereby facilitating exploration of how to best represent complex spatial variations of ground-motion amplitude in GMPEs.

Presenting Author: Brad Aagaard

Authors