| ||Format||Pages||Price|| |
|PDF (280K)||17||$25||  ADD TO CART|
|Complete Source PDF (7.1M)||428||$67||  ADD TO CART|
A three-layer, feed-forward artificial neural network was trained to perform a tomographic inversion of seismic travel time data for a simple 4 by 4 grid. One hundred training sets were created by randomly varying the velocity of one of the four interior cells. Each training set consisted of the travel times for 33 ray paths traversing the grid from sources and receivers placed on three sides. Ray curvature was accounted for in the travel time calculations. An additional 100 sets of data were used to test the performance of the trained network. The network performed well when the velocity of anomalous cell was larger than the background velocity or slightly less. For anomalous cells with velocities significantly lower than the background velocity the network was incapable of providing correct results because of refraction around the low-velocity cell. The trained network successfully generalized for several test cases involving two anomalous cells and was reasonably robust when presented with noisy travel time data.
tomography, artificial neural network, inversion, refraction, ray paths, seismic, geophysical
Assistant Professor, School of Civil Engineering, Georgia Institute of Technology, Atlanta, GA