HR: 0800h
AN: NG31B-0873    [Abstracts]
TI: Support Vector Machines for Non-linear Geophysical Inversion
AU: * Kuzma, H A
EM: andheidi@uclink.berkeley.edu
AF: University of California Berkeley, P.O. Box 987, Truckee, CA 96160
AU: Rector, J W
EM: jwrector@lbl.gov
AF: University of California Berkeley, P.O. Box 987, Truckee, CA 96160
AB: Classical non-linear geophysical inversion can be simulated using computer learning via Support Vector Machines. Geophysical inverse problems are almost always ill-posed which means that many different models (i.e. descriptions of the earth) can be found to explain a given noisy or incomplete data set. Regularization and constraints encourage inversions to find physically realistic models. The set of preferred models needs to be defined a priori using as much geologic knowledge as is available. In inversion, it is assumed that data and a forward modeling process is known. The goal is to solve for a model. In the SVM paradigm, a series of models and associated data are known. The goal is to solve for a reverse modeling process. Starting with a series of initial models assembled using all available geologic information, synthetic data is created using the most realistic forward modeling program available. With the synthetic data as inputs and the known models as outputs, a Support Vector Machine is trained to approximate a local inverse to the forward modeling program. The advantages of this approach are that it is honest about the need to establish, a priori, the kinds of models that are reasonable in a particular field situation. There is no need to adjust the forward process to accommodate inversion, because SVMs can be easily modified to capture complicated, non-linear relationships. SVMs are transparent and require very little programming. If an SVM is trained using model/data pairs that are drawn from the same probability distribution that is implicit in the regularization of an inversion, then it will get very similar results to the inversion. Because SVMs can interpret as much data as desired so long as the conditions of an experiment do not change, they can be used to perform otherwise computationally expensive procedures. Support Vector Machines are trained to emulate non-linear seismic Amplitude Variation with Offset (AVO) inversions, gravity inversions and electromagnetic inversions. Training an SVM, including generating training data, is generally much faster than performing a non-linear inversion.
DE: 9810 New fields (not classifiable under other headings)
DE: 9820 Techniques applicable in three or more fields
DE: 3299 General or miscellaneous
DE: 0699 General or miscellaneous
DE: 0900 EXPLORATION GEOPHYSICS
SC: Nonlinear Geophysics [NG]
MN: 2004 AGU Fall Meeting