Principal Investigator: Mrinal K. Sen

Funded by: National Science Foundation

The purpose of this project is to develop and implement inversion methods based on neural networks for application to geophysical inversion problems, particularly, seismic inversion. The motivation for investigating this approach is as follows: for geological interpretation, the mapping of elastic properties (impedance, Poisson's ratio, etc.) along a seismic line is required. GA inversion for each point along a seismic line is currently computationally expensive. However, for a selected number of points along a line Genetic Algorithms to derive the elastic properties can be used. During that process a neural network can be trained, i.e., derive the connection weights. After training, seismic gathers from the intermediate points can be input to the network to estimate depth-depedent elastic properties for those surface points resulting in a complete map. The most difficult part of the problem is to train the network. Gobal optimization methods such as Simulated Annealing and Genetic Algorithms are used for the network training since the methods have been proven robust in application to geophysical inversion.