Predicting Log Data by Using Artificial Neural Networks to Approximate Petrophysical Parameters of Formation

Abstract

Petrophysical characteristics of underlying formation have an important role in reservoir management and drilling wells. One of the most common ways to reach this information is well log analysis. These logs represent important parameters of formation by evaluating some characteristics of rocks. Meanwhile, some time well logging does not implement well or some log data are accompanied by so many errors. Therefore, in these cases using cheaper logs or better data can estimate more expensive log or incorrect data. High-skill experts and lithology information are needed for interpretation and evaluation of data. Therefore, designing a model that is able to evaluate the petrophysical index using well log data without laboratory information will be very economical. The authors solve the problem of well log data and interpretation of logs using artificial neural networks. Running various networks has implied that the selection of appropriate network's input is the most important factor in accurate estimates. After selecting the appropriate input data, many networks were run in order to optimize the number of epoch, hidden layers, neurons in each layer and the proper neural network's training function. First, sonic log (DT) is predicted using neural network system. To do this, neutron porosity (NPHI) and density log (RHOB) are used as input variables. In the second network, porosity index (PHIE) is evaluated and predicted considering NPHI, RHOB, and DT. Finally, in third network, saturation index is estimated by using resistivity log and PHIE is predicted by second network.