Model parameter optimization for site-dependent simulation of ground motion by simulated annealing: Re-evaluation of the Ashigara Valley prediction experiment

More than 40 groups from 10 different countries participated in a weak- and strong-motion prediction experiment at Ashigara Valley which required the blind prediction of time series, spectra and spectral ratios for selected and instrumented sedimentary sites with well-known geotechnical properties. The wide scatter of the results of this experiment have raised a number of questions as to how to model high-frequency ground motion in the presence of available geotechnical and geophysical data. Using a simulated annealing waveform inversion method, we have tried to optimize and automate the model construction for ID site-dependent ground-motion simulation. We found a whole set of successfull models which provide good waveform fit (r > 0.8) for the observed displacement records at site KS2 but also yield sufficiently accurate response spectra and peak value predictions for both surface and downhole site. This shows that ID models are fully adequate to model the site conditions at least for the weak motion data. The resulting successful layer models consistently show a slower, less dense, and slightly thicker low velocity coverage with stronger damping than the official geotechnical model. Furthermore, their statistical properties directly measure the sensitivity of the individual parameters for the simulations. The critical re-evaluation of our own prediction which was based on stochastic simulation shows that although this approach has its greatest merits in situations where little information is available, it can also be successfully applied to model individual records if sufficient care is taken to determine the source parameters. Simulated annealing waveform inversion has shown to be a powerful tool to optimize that process.