Effect of lithological variations on the performance of artificial intelligence techniques for estimating total organic carbon through well logs

By the expansion of production from source-related unconventional petroleum resources, accurate approximation of Total Organic Carbon (TOC) through well logs has become progressively important. Accordingly, recent studies have mainly focused on increasing the precision of TOC estimation by using different types of AI techniques and/or optimizing algorithms. Along with utilizing these approaches, this study emphasized on removing an unaddressed source of error inherited from lithological heterogeneity with the same goal. Like organic matter quantity, lithological variations within a source interval also induce well log responses, which may interfere with the training process of Artificial Intelligence (AI) techniques. In the present research, the effect of lithological variations on the performance of TOC estimators was evaluated by employing Adaptive Neuro Fuzzy Inference System (ANFIS) and Multilayer Perceptron network (MLP). Firstly, ANFIS and MLP models were constructed and trained using a database containing different lithologies (original models). Then, a new methodology was developed based on modeling the relationship between log data and TOC values for each type of lithology (litho-based method). The result showed that the litho-based method estimates more reliable and accurate TOC values, proving the adverse effect of lithological variations on the original models. Furthermore, the litho-based ANFIS models provide the most promising results. Since metaheuristic algorithms are usually employed to optimize AI techniques, Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) were also implemented into the original models (hybrid models). Accuracy of TOC values estimated by the hybrid models is slightly higher than those derived from the original models. However, these hybrid approaches are not as efficient as the litho-based method. Applicability of the proposed approach was guaranteed by performing it over Pabdeh source rocks in a well of SW Iran. Based on its high efficiency, the newly developed litho-based method can be used as a powerful tool to reliably evaluate unconventional hydrocarbon resources, as well as the potential of the conventional petroleum sources. Moreover, it can be utilized, instead of/along with traditional optimization approaches, to approximate other geochemical factors as well as petrophysical parameters from log data.