Doctoral Dissertation 2004

Two alternative inversion techniques for the determination of seismic site response and propagation-path velocity structure: spectral inversion with reference events and neural networks

Cesar Aaron Moya

This work corresponds to a two-part study in the Tottori region for which site effects and velocity structure were analyzed using non-conventional techniques. Part one of the study presents a method to determine site effects and Q factor using a reference event. Spectral ratios between pairs of earthquakes and a reference one were written as logarithmic summations in order to obtain individual solutions for the numerators and denominators. Given the seismic moment and corner frequency for one of the events following the omega-squared model, a source model was assumed to constrain the amplitude of the solution. The validity and limitations of this technique were tested using synthetic data. We analyzed the variances of the results depending on the different number of stations used, the amount of noise present in the records, and the change in the seismic moment, corner frequency, and source model of the reference event. Aftershocks of the 2000 Tottori, Japan, earthquake were used to estimate the site effects and Q factor in that region. Two strong motion networks recorded the aftershocks: K-NET and KiK-net (KiK-net contained surface and borehole receivers at every station). Data were inverted first using only surface records and then using the borehole records from KiK-net stations. Our results suggest that the study area presents a low Q value and that there is also amplification at borehole sites.

Part two is about a velocity structure inversion approach using neural networks (NN). Four events from the Tottori sequence were selected around station SMNH01 in order to determine a 1D nearby underground velocity structure. A NN was trained for each earthquake-station profile using synthetic data. Upon training, actual observed records of the four events were given as input to the network which tried to predict their corresponding velocity structure. First, simple 1D profiles were obtained individually for each of the events. Then, the validity of each model was tested by analyzing the waveform fitting of different events recorded at SMNH01 and two other nearby stations: TTR007 and TTR009. We also analyzed a 3D case in which the depth and P and S-wave velocities of a basin-like structure were inverted using NN. S-wave velocity is given assuming a constant Poisson ratio. For this test, only the upper layer of the model was inverted using synthetic data. In general, the network was able to predict most of the models accurately. It was observed that as many as 75% of the models were estimated with a goodness-of-fit of 70%.