Machine learnig aplicado à caracterização petrofísica de reservatórios: um exemplo da bacia do Ceará

Abstract

In the process of reservoir characterization and exploration, geophysical well logging stands as a widely employed tool capable of estimating various physical properties of the explored environment. It also evaluates the content of existing fluids, thus providing an estimate of recoverable oil volume. In this context, machine learning techniques can be integrated as tools to predict such properties, thereby enhancing petrophysical evaluation workflows and enabling a deeper comprehension of log data. In this regard, the present study aims to employ well logging data for reservoir characterization through the development of both conventional and machine learning-based workflows. Initially, well 3-EP-5-CES from the Espada field in the Ceará Basin was selected for the identification and assessment of contained reservoirs. Various petrophysical calculations were conducted to estimate its properties. The proposed conventional petrophysical analysis identified 4 zones for economic evaluation, yielding average values of 20% for clay content, 30.28% for porosity (φ), 67.65% for water saturation (So), and 643 mD for permeability. This resulted in a Net-Pay of 14m, aligning closely with values commonly described for the studied field as per the ANP executive summary. Furthermore, the study classified three electrofacies - reservoir, potential reservoir, and non-reservoir - using an unsupervised clustering technique, k-means, based on well 3-EP-5-CES log data. This approach automated the reservoir facies identification and obtained calculated petrophysical parameters of 20.79% for clay content, 27.65% for φ, and 53.66% for So, akin to results obtained through conventional evaluation. Additionally, the methodology employing neural networks for predicting lithological facies, Sw curves, and φ of well 3-CES-38D-CE demonstrated a good prediction accuracy compared to actual data. This underscores the effectiveness of the adopted model as an estimation approach. The lithology prediction proved efficient, particularly for sandstones, achieving an accuracy of 72.02%, offering a viable solution when lithological information is lacking. The results of Sw curve prediction optimized the conventional workflow, maintaining resemblance to conventional data with an average error of 0.1693. Lastly, NPHI prediction also showcased promise in scenarios with missing data, with mean errors below 0.18.