The magnetotelluric (MT) method is a natural source electromagnetic geophysical technique, which is used mainly in petroleum, mineral and geothermal exploration. As in this method, the quantity of the measured data is bulky and have a complex structure, their modeling, compared with the modeling of the other electrical data, is a very complex task or even impossible in some instances.
The main objective of this paper is to use the ability of the artificial neural networks (ANN) to find a solution for two-dimensional (2D) joint TE (transverse electric) and TM (transverse magnetic) modes inverse modeling of MT data. To achieve the goal, a multilayer perceptron (MLP) network with back propagation (BP) learning algorithm is used. In order to learn the designed network, many synthetic 2D models with the same category, have been created and their responses have been calculated for each polarization mode by forward modeling. Synthetic data include apparent resistivity and impedance phase in 9 stations and 11 frequencies in two polarization modes. After a comprehensive study, a perceptron with 3 layers and architecture of 396-9-9 has been designed and used to model the data.
This study show that the designed network is capable enough to produce an acceptable 2D underground model so that the correspondence mean relative modeling error is 3.9% and 6.9 % respectively for noise free data and 5 percent randomly added noisy data. This indicates that if ANN is designed and trained properly, then it would be capable enough to perform 2D inverse modeling of MT data. It has also shown that once the designed network has been trained properly it is able to perform the inverse modeling precisely in a short time. At the end, the performance of the designed network has been evaluated by a set of field MT data and its results has been compared with those produced by a common smooth rapid relaxation inversion (RRI) method. The comparison indicates that the results of these two different procedures are in close agreement.