Inversion of self-potential data using generalized regression neural network

This paper presents a method for parameter estimation of self-potential (SP) anomalies using neural networks. The General Regression Neural Network (GRNN) one-pass learning algorithm was performed to invert SP anomalies of simple shaped geometrical bodies approximation. The one-pass learning algorithm has a certain advantage in terms of computation time compared to classical neural networks because the classical neural networks use multiple learning steps. The presented algorithm was tested on noise-free and noise-corrupted synthetic data. In addition, the method was applied to three field examples: Süleymanköy, Weiss, and Sarıyer anomalies, respectively. The model parameters including electric dipole moment, polarization angle, depth, shape factor, distance from the origin of the anomaly, base slope and the base level were successfully estimated using the presented method. The frequency distribution of each model parameter was calculated to improve and overcome the ambiguity of the estimated model parameters. To investigate the correctness of the estimated model parameters, the obtained results were compared with previous studies. Thus, the agreement between the results obtained by the present method and other previous results is similar to most of the estimated model parameters in accordance with numerical values. The result of the present study shows that the GRNN can be used as a powerful parameter estimation tool in the interpretation of SP data in terms of computation time compared to artificial neural networks.