Permeability estimates of a carbonate reservoir in Campos Basin, Southeastern Brazil, using well logs with empirical, multilinear regression and machine learning approaches

Abstract

Well logging records the physical properties of geological formations and the fluids traversed by the wells. This operation is interested in parameters such as lithology, hydrocarbon presence, permeability, porosity, and fluid saturation. Generally, oil reservoirs are sandstone or carbonate rocks, and the latter’s characterization is a critical question in the petrophysical properties distribution, mainly permeability. Estimating permeability is a complex task due to the heterogeneity of these reservoirs. Therefore, this work used conventional logs to estimate the permeability of wells A03 and A10, both belonging to the oilfield A, Campos Basin, Southeastern Brazil. Alongside the logs, permeability measured in the laboratory in rock samples was used to validate the achieved estimates. Thus, the estimates used basic logs as input and approaches such as Timur empirical equation, multilinear regression, and machine learning techniques, like fuzzy logic, neural network, and decision tree. Pearson's coefficient of determination R was used as the comparison metric with the experimental data. The number of samples in training was 70%, with 15% in the validation and testing, and the results show that the first four estimates showed bad fits (R£0.60), while the decision tree showed good fits (R>0.60). This approach also showed that the gamma-ray and resistivity logs are the ones that have the most significant weight in the estimates.