Machine learning applications for core guided petrophysical analysis

Abstract

Rock physics models are used to characterize and quantify rock properties such as reservoir quality or fluid saturation. The factors that control these rock properties are complex, and they require numerous assumptions and approximations. Good data quality is essential for reliable and consistent interpretation. In this thesis, I present two methods for gas hydrate saturation estimation using well log and core data, namely the resistivity method, and the nuclear magnetic resonance (NMR) - density method. Then, I show the discrepancy between the two models to highlight the data quality issues. After that, I present an analysis of several supervised learning algorithms to produce synthetic density and porosity logs using pressure-core data as ground truth and ultimately improve saturation estimation. Also, I apply an unsupervised learning algorithm to correlate the saturation estimation values from the resistivity method with different clusters of the raw NMR T2 spectrum. I compared the theories, underlining assumptions, strengths, and weaknesses of the two saturation estimation methods. After that, I apply a forward modeling technique for estimating the velocity effect of gas hydrate using the saturation estimation curves. Then, I discuss the supervised machine learning workflow from data preparation, feature selection, model selection, and model evaluations. Four supervised machine learning algorithms were selected for model evaluation, namely linear regression (LR), random forest regressor (RF), multilayer perceptron (MLP), and long short-term memory (LSTM). For the unsupervised learning, the K-mean clustering algorithm was applied to classify NMR T2 spectra and correlated with our saturation estimation values. My results indicated that the resistivity method seems to be slightly better than the NMR-density method; it produced lower misfits between the measured and the modeled velocities and therefore considered more representative for estimating saturation. The machine learning analysis showed that the deep neural network LSTM model had the best overall performance. The LSTM model generated density and porosity logs which reduced the mean error by 50% and produced a more reliable saturation analysis. The unsupervised learning result showed that there is a correlation between the estimated gas saturation value from the resistivity log and the NMR T2 amplitude signal, which suggests the raw NMR T2 spectra are sensitive to the presence of free gas.