Lithofacies recognition based on fuzzy logic and neural networks: A methodological comparison

Abstract

Fuzzy logic and neural networks methods are commonly applied in various areas of the petroleum industry. In this work, both methods are used comparatively to lithofacies prediction in wells. The input dataset available for the study includes 4 known well logs (gamma ray, neutronic porosity, Density and Sonic) and lithologic description of cores for 14 wells drilled in the Namorado oil field. This field is located in the Campos Basin (SE Brazil) and its reservoir rocks are composed of sandstone turbidites. Core descriptions correspond to 3,196 samples of facies distributed in irregularly spaced intervals of 14 wells. Among the samples, 21 different facies were recognized by the geologists. These lithofacies were regrouped into three representative lithofacies: sandstone, shale and limestone. Fuzzy logic and backpropagation neural network models were produced using data from three key wells. Each model was applied to all other 11 wells. The comparison was individually performed between the original core intervals and the synthetic lithofacies column. Results revealed that for the original 3,196 facies samples, 2,353 were correctly recognized by the fuzzy logic method, whereas 2,599 were correctly predicted by neural networks. These correspond to approximately 73% and 83% accuracy, respectively. The neural networks method also exhibited enhanced results for each well separately. In summary, both methods performed well in recognizing three main lithofacies intervals for 14 wells at the Namorado Field. The analysis indicates that the methods showed good performance in the recognition of lithofacies for all wells investigated in the Namorado oil field. In general, neural networks showed a product with accuracy around 10% higher than that obtained by fuzzy logic. The accuracy yielded by neural networks is also higher when wells are compared individually. The superiority of the results obtained with neural networks suggests better ability of this algorithm on the recognition of lithofacies, particularly in geologic scenarios similar to those approached here.