HR: 12:05h
AN: S42A-07 [Abstracts]
TI: Using Borehole Vertical Array Data to Determine Local Attenuation and Velocity Structure: A Combined
Global-Local Optimization Algorithm for Plane Wave Seismogram Inversion
AU: Assimaki, D
EM: dominic@crustal.ucsb.edu
AF: University of California at Santa Barbara, Institute for Crustal Studies
1140 Girvetz Hall, Santa Barbara, CA 93106-1100
United States
AU: Tsuda, K
EM: kenichi@crustal.ucsb.edu
AF: University of California at Santa Barbara, Institute for Crustal Studies
1140 Girvetz Hall, Santa Barbara, CA 93106-1100
United States
AU: Oakes, J
EM: joakes@crustal.ucsb.edu
AF: University of California at Santa Barbara, Institute for Crustal Studies
1140 Girvetz Hall, Santa Barbara, CA 93106-1100
United States
AU: * Steidl, J
EM: steidl@crustal.ucsb.edu
AF: University of California at Santa Barbara, Institute for Crustal Studies
1140 Girvetz Hall, Santa Barbara, CA 93106-1100
United States
AB:
A seismic waveform inversion algorithm is demonstrated for the estimation of elastic soil properties from one-dimensional
downhole array recordings. For a given bedrock motion, scarcity of near-surface geotechnical information, error propagation
and limited resolution of the continuum usually result in predictions of surface ground motion that poorly compare with low
amplitude observations. This discrepancy is further aggravated for strong ground motion, associated with hysteretic,
nonlinear, and potentially irreversible material deformations. Seismogram inversion is a nonlinear multi-parameter
optimization problem. Traditional search techniques that use characteristics of the problem to determine the next sampling
point (e.g. gradients, Hessians, linearity and continuity) are computationally efficient, yet limited to convex regular
functions. As a result, they fail to identify the best fit solution in seismogram inversion problems, when the starting model
is too far from the global optimal solution. On the other hand, stochastic search techniques (e.g. genetic algorithms,
simulated annealing) have been shown to efficiently identify promising regions in the search space, but perform very poorly
in a localized search.
The proposed inversion technique is a two-step process, namely a genetic algorithm in the wavelet domain in series with a
nonlinear least-square fit in the frequency domain; we thus improve the computational efficiency of the former, while
avoiding the pitfalls of using local linearization techniques such as the latter for the optimization of multi-modal,
discontinuous and non-differentiable functions. The parameters to be estimated are stepwise variations of the shear modulus,
attenuation and density with depth, for horizontally layered media with refined near-surface discretization. Equality
constrains are imposed on the vector of unknowns to bound the search space, based on the available soil investigation. For
the genetic algorithm, the objective function is defined as the normalized cross-correlation between the observed data and
the synthetics. We perform the inversion in the wavelet domain to allow for equal weighting of the information across all
frequency bands. Since ground motion is non-stationary in time and frequency, a time-domain inversion would inevitably
emphasize the larger amplitude signals. The process is repeated in series for a subset of the available borehole and surface
waveform pairs, selected on the basis of signal quality. The mean estimated soil properties from the genetic algorithm are
then used as a starting model for the local minimization scheme. The target function in this stage is the empirical transfer
function in the frequency domain, estimated using the average spectral ratio between surface and borehole pairs.
The global-local inversion technique can efficiently identify the optimal solution vicinity in the search space by means of
the hybrid genetic algorithm, whereas the use of nonlinear least-square fit accelerates substantially the detection of the
best fit model. The algorithm has been implemented in MATLAB, and inversion results are illustrated for stations in the
Japanese strong motion borehole array Kik-Net, as well as for borehole stations in Southern California jointly operated by
the California Integrated Seismic Network, the Southern California Earthquake Center, and the University of California at
Santa Barbara.
DE: 7223 Seismic hazard assessment and prediction
DE: 7212 Earthquake ground motions and engineering
SC: Seismology [S]
MN: 2004 AGU Fall Meeting