Caracterização de petrofácies e genese das rochas carbonáticas do reservatório Sapinhoá (Pré-sal, Bacia de Santos) e aplicação de aprendizagem de máquina por meio de microscopia óptica e imagens digitais a partir de atributos permo-porosos

Abstract

The central focus of this doctoral thesis comprises the characterization of pre-salt carbonate reservoirs and the development of an innovative approach to predict the porosity and permeability of these rocks. This prediction method utilizes machine learning techniques, eliminating the dependency on geophysical data. The study is structured into three distinct manuscripts: (I) Petrography and Genesis of Carbonate Rocks from the Barra Velha Formation (Santos Basin, Brazil) and Associated Biogenic Structures, (II) Prediction of Porosity and Permeability of Carbonate Rocks Using Thin Section Data and Machine Learning Methods: A Systematic Review, and (III) Prediction of Carbonate Rock Porosity and Permeability Properties Using Convolutional Neural Networks. The techniques employed encompassed a conceptual literature review, as well as methodologies in machine learning, optical petrography of carbonate rocks, and convolutional neural networks. In the initial phase, 565 thin sections of pre-salt rocks were described through optical petrography, obtained from three wells within the Santos Basin. These descriptions facilitated the identification of primary carbonate facies, elucidated their genesis, and provided the dataset essential for training the neural network. Subsequently, a systematic literature review of prominent machine learning techniques was conducted, integrating the datasets obtained in the previous phase. The primary aim was to leverage these datasets for predicting porosity and permeability. Lastly, in the third phase, the training and prediction of porosity and permeability of pre-salt carbonate rocks were executed utilizing the same dataset utilized in the initial phase. Optical petrography enabled the identification of distinct carbonate facies within the pre-salt rocks, alongside their textural characteristics, mineralogy, and void percentage within the rock matrix. These attributes collectively formed the foundational components of the training dataset for poroperm prediction. Moreover, insights can be discussed regarding the genesis of pre-salt formations, contributing to the evaluation of the complexity inherent in these substantial oil and gas reserves. Concurrently, systematic bibliographic analysis aids in comprehending the most prevalent techniques employed for dataset generation. Ultimately, the training phase facilitated the development of an specific architecture for the dataset and the particular purpose of predicting porosities and permeabilities of pre-salt rocks, without the necessity for supplementary geophysical data. The integration of data from the optical petrography phase not only facilitated the acquisition and selection of the training dataset but also played a pivotal role in analyzing the data obtained during the predictive process. This phase encompassed the treatment and selection of the image dataset. Moreover, exploration of hydrothermal and biogenic evidence yields valuable insights into the genesis of carbonate facies and their impact on reservoir porous properties. The results obtained from the petrographic analysis of wells in the Santos Basin corroborate with the understanding of the importance of formation processes in determining the porous properties of rocks. The utilization of a convolutional neural network architecture enabled the extraction of porosity and permeability values from a heterogeneous and complex dataset, underscoring the efficacy of the developed architecture. It points out high predictive capacity for both porosity (R2 = 0.86) and permeability (R2 = 0.96), establishing itself as a fundamental tool for characterizing reservoirs in pre-salt carbonate environments. Subsequently, further experimentation will be conducted employing an expanded dataset comprising a larger number of images, alongside a novel hybrid approach, integrating images and labels obtained from the optical petrography phase for training purposes.