Geophysical Modeling Using Neural Network

@inproceedings{Foudil-BeyRM2008,
 abstract = { This work aims at determining 3D distribution model parameters such as density, susceptibility, and conductivity from geophysical surface survey responses (gravity, magnetic, and electromagnetic maps) using a Neural Network approach. It is usual for geophysics airborne surveys to cover very large areas extending over several kilometers squares. Therefore, the computer time requested to invert these geophysical data may takes several weeks depending on the model resolution and on the power of the machine which perform the inversion. Once the learning part is finished and the network has been set up, the main advantage of Neural Networks in potential field modeling is its ability and rapidity for determining the model parameters. In a first step, the network is trained with synthetic examples in order to reduce the differences between the predicted input and ideal theoretical output, basically these differences decrease with iterations number. The back-propagation techniques are used for adjusting the network parameters at each iteration step, and finally, the training process is stopped after an a priori user fixed number of iterations. Then, the final fit is validated by comparing the predicted output by the network with the ideal output. This method has been implemented as a gOcad plug-in and tested on gravity data sets. Firstly, the algorithm computes the gravity field from different models after using the learning process in order to find out the proper network parameters. Then, the procedure consists in determining the output of the network (prediction of the model parameters) from a given input (gravity field). This approach is applied into a case study. },
 author = { Foudil-Bey, Nacim AND Royer, Jean-Jacques AND Cheng, Li Zhen },
 booktitle = { 28th gOcad Meeting },
 month = { "june" },
 publisher = { ASGA },
 title = { Geophysical Modeling Using Neural Network },
 year = { 2008 }
}