Machine Learning-Assisted Prediction of Electrical Properties for Heterogeneous Reservoirs: Case study in Mamey Field, Northern Colombia

Abstract

Abstract Petrophysical characterization in reservoirs with high heterogeneity is always a challenge and the application of state-of-the-art technologies plays an increasing role in achieving accurate and timely results. The target reservoir is composed of interlaminated mudstones and very fine to fine sandstones enclosed in a deltaic environment and capped by cross-stratification sandstones associated with incised valley deposits. The depositional environment creates the geological and petrophysical complexity of the field. This case study describes an automatic learning technique or "Machine Learning" to determine the electrical properties "m" and "n". The methodology combines logs, rock types, facies and digital core analyses from Mamey field, Lower Magdalena Valley basin, northern Colombia. An artificial neural network (ANN) methodology based on a self-organizing map (SOM) was applied. The ANN was trained via unsupervised learning to produce a low-dimensional discrete representation of the input space of the training samples called a map. The method does achieve a reduction in dimensionality and creates a continuous curve. An SOM output is typically a branch based on the cluster. The results obtained indicate that the machine learning technique is feasible to estimate a continuous curve of the Archie parameters "m" and "n" associated with the textural changes identified in the images and computed tomography. The variables "m" and "n" strengthened the resistivity-dependent water saturation model that affects in-situ gas and hence reserves calculations. In addition, the constructed rock electrical properties model reduces costs associated with special core analyses, especially in highly heterogeneous and complex formations.