Machine learning approaches for petrographic classification of carbonate-siliciclastic rocks using well logs and textural information

The definition of lithology in oil wells by means of multiple geophysical analysis profiles an important role in the reservoir characterization process. From this one can generate lithologic models which in turn will be filled in with the petrophysical properties and can then be used in flow simulators to understand and study the behavior of an oil field. The identification of lithology can be accomplished by direct and indirect methods, but these procedures are not always feasible because of the cost or imprecision of the results generated. Consequently, there is a need to automate the process of reservoir characterization and, in this context, computational intelligence techniques appear as an alternative to lithology identification. In this paper balancing strategies are implemented in order to tackle the lack of data due to the characteristics of the distinct diagenetic processes. Six machine learning methods combined with data balancing techniques and a model selection approach to classify geologic data from the South Provence Basin. The results show the balancing strategies improve the overall performance of all classifiers and the model selection allows or obtaining the best parameters from a user-defined set. The computational tool developed in this paper arises as an alternative to assist geologists ans specialists in the task of reservoir heterogeneities identification.