Depth and shape estimation of salt domes via interpretation of gravity data using Multi Layer Perceptron Neural Networks

In applied geophysics especially in potential methods like gravity generalized bodies are often used to represent the distribution of underground masses, as sphere, vertical cylinders, vertical prisms, horizontal cylinders, vertical faults, anticlines and synclines. In this paper Multi Layer Perceptron (MLP) Artificial Neural Networks are used to find the most probable model for a given gravity anomaly of a salt dome. Therefore a neural network is trained with anomalies produced by two different kinds of distributing bodies, producing similar anomalies. These simple models which are the most common used shapes for modelling of salt domes are Sphere and Vertical Cylinder. The trained Multi Layer Percetron Artificial Neural Network is then able to recognize the kind of body that is producing the given gravity anomaly. Throught neural networks technique the ambiguity between similar anomalies generated by different disturbing bodies can be solved without using densities. There is no classical interpretation method available, which can, for example discriminate between an anticline and a syncline without any hypotheses about the shape or density contrast of the target.It is shown here that this can be done by applying Multi layer perceptron Artificial Neural Networks for qualitative gravity interpretation. By using of this kind of Artificial Neural Networks the gravity data interpreter can do qualitative and gravity quantitative interpretation. Qualitative interpretation means to solve the ambiguity between two bodies that produce similar anomalies. In quantitative interpretation with multi layer perceptron Artificial Neural Networks, the model parameters (include depth, radius) can be achived. Sphere and vertical cylinder are the models to representing the salt domes. Therefore, as we use data gravity of humble salt dome, as a real test of the method, we will use these models for training of the neural network. By using of sphere and vertical cylinder models, we prepared, normalized and used a set of suitable features as inputs of the network. Because there is no certain rule for defining the suitable number of the neurons of hidden layer, by changing the number of neurons in hidden layer, and comparing the Sum Squared Errors in every state, we received best number for neurons for this layer. After defining these neurons, by synthetic data from artificial sphere and cylinder models, trained the network. It is necessary to mention that the neural network was trained in the relatred domain of thre probable depth, especially for the real data that we know the geological prior information and so the approximation of the depth domain is possible.Also the training data are all normalized both inputs and outputs. The index used to evaluate the errors was sum squared error for both validation and test data. Finally by using outputs of the network used for recognition of the shape of the anomaly, and the network used for defining model parameters, we defined the shape and parameters of humble salt dome. The results for real and synthetic gravity data showed very good ability of the multi layer perceptron neural networks for estimation of shape and depth of salt domes.