multi-layer perceptron

Full identification and understanding of hydrocarbon reservoirs depends on knowing the mechanical properties. One of the main parameters that indicates mechanical properties is shear wave velocity. Bipolar sound recorder is among thee best tools for measuring shear wave velocity. This tool is not very popular due to the high costs of driving in the well despite the high accuracy. Shear wave velocity estimation methods include three main branches of experimental methods, regression and the use of machine learning algorithms or in other words artificial neural networks. The studied formation in this research is Sarvak in one of the oil fields in the south of Iran. The input data of the estimator model is the usual petrophysical logs that are driven and measured in many wells, and the output data is obtained from the DSI tool. In this research, data are pre-processed by removing noise effects. Then, to improve the estimation effectiveness, data with a high correlation coefficient are selected as input data. After that, shear wave velocity is estimated from petrophysical data with three types of multi-layer perceptron (MLP), multi-layer perceptron optimized by particle swarm optimization (MLP-PSO), and the introduction of a relatively new method of multi-layer perceptron-social ski drive (MLP-SSD). To compare the efficiency of the neural network method, two traditional experimental and regression methods used. The validation results show the better performance of the MLP-SSD method.

The purpose of this study is to calculate Total Organic Carbon (TOC) values of the Iranian field using a combination of sonic and resistivity logs (Passay method) and neural networks method in the conditions, where the core analysis or well-log measurement does not exist. We compared the resultant TOC with the ones obtained from the geochemical analysis. To correlate between the total organic carbon data and petrophysical log, which are available after logging, Multilayer Perceptron Artificial Neural Network is used. After analyzing 100 cutting samples by using rock -Eval pyrolysis, geochemical parameters have achieved.By using the multi-layer perceptron with Levenberg–Marquardt training algorithm, the TOC with correlation coefficient 0.88 and MSE 1.443 have been provided in the intervals without analyzed samples. Finally, the TOC was estimated by using separation of resistivity and the sonic log, although, with the favorable results in some other fields, the estimation had a correlation coefficient of 51% in this field. Comparing the performance of the multi-layer perceptron with Levenberg–Marquardt training algorithm (with an accuracy of 88%) and results of the Passay method (with an accuracy of 51%) indicated that the neural network is more accurate and has better consistency compared with the empirical formula.