S41C-01
How Much Can the Total Aleatory Variability of Empirical Ground Motion Prediction Equations Be Reduced Using Physics-Based Earthquake Simulations?
Thursday, 17 December 2015: 08:00
307 (Moscone South)
Thomas H Jordan1, Feng Wang2, Robert W Graves3, Scott Callaghan4, Kim Bak Olsen5, Yifeng Cui6, Kevin R Milner4, Gideon Juve7, Karan Vahi7, John Yu4, Ewa Deelman7, David Gill4 and Philip J Maechling1, (1)Southern California Earthquake Center, Los Angeles, CA, United States, (2)AIR-Worldwide Corporation, Boston, MA, United States, (3)USGS California Water Science Center Menlo Park, Menlo Park, CA, United States, (4)University of Southern California, Los Angeles, CA, United States, (5)San Diego State Univ, San Diego, CA, United States, (6)San Diego Supercomputer Center, La Jolla, CA, United States, (7)USC Information Sciences Inst., Marina Del Rey, CA, United States
Abstract:
Ground motion prediction equations (GMPEs) in common use predict the logarithmic intensity of ground shaking, lnY, as a deterministic value, lnYpred(x), conditioned on a set of explanatory variables x plus a normally distributed random variable with a standard deviation σT. The latter accounts for the unexplained variability in the ground motion data used to calibrate the GMPE and is typically 0.5-0.7 in natural log units. Reducing this residual or “aleatory” variability is a high priority for seismic hazard analysis, because the probabilities of exceedance at high Y values go up rapidly with σT. adding costs to the seismic design of critical facilities to account for the prediction uncertainty. However, attempts to decrease σT by incorporating more explanatory variables to the GMPEs have been largely unsuccessful (e.g., Strasser et al., SRL, 2009). An alternative is to employ physics-based earthquake simulations that properly account for source directivity, basin effects, directivity-basin coupling, and other 3D complexities. We have explored the theoretical limits of this approach through an analysis of large (> 108) ensembles of 3D synthetic seismograms generated for the Los Angeles region by SCEC’s CyberShake project using the new tool of averaging-based factorization (ABF, Wang & Jordan, BSSA, 2014). The residual variance obtained by applying GMPEs to the CyberShake dataset matches the frequency-dependence of σT obtained for the GMPE calibration dataset. The ABF analysis allows us to partition this variance into uncorrelated components representing source, path, and site effects. We show that simulations can potentially reduce σT by about one-third, which could lower the exceedance probabilities for high hazard levels at fixed x by orders of magnitude. Realizing this gain in forecasting probability would have a broad impact on risk-reduction strategies, especially for critical facilities such as large dams, nuclear power plants, and energy transportation networks.
