Application of Machine Learning for Estimating Petrophysical Properties of Carbonate Rocks Using NMR Core Measurements

Abstract

Abstract Evaluation of petrophysical properties such as porosity, permeability, and irreducible water saturation is crucial for reservoir characterization to determine the hydrocarbon initially in place and further optimize hydrocarbon production. However, estimation of these parameters is challenging for carbonate rocks due to their heterogeneity. One of the ways to determine petrophysical properties is the use of nuclear magnetic resonance (NMR), which involves applying a magnetic field to the formation and detecting signals emitted from pore spaces. The main objective of this study is to develop an empirical correlation for porosity, permeability, and irreducible water saturation by comparing NMR and laboratory measurements for carbonate rocks in the Middle East. Furthermore, machine learning (ML) approach was applied to predict these petrophysical parameters utilizing NMR data. Different ML algorithms such as tree-based and neural networks were trained to estimate these petrophysical properties of carbonate rocks. The obtained results from ML algorithms were further compared with core measurements to ensure their accuracy. The results showed that the use of T2 spectrum as an input provided more accurate results than NMR features. It can be proven by observing the performance of deep neural networks algorithm, where the models showed R2 values of 0.87 and 0.74 for porosity prediction using T2 and features extraction approaches, respectively. The same behavior was followed for the permeability estimations as deep neural networks model scored R2 = 0.81 (T2 approach) and R2 = 0.74 (features extraction approach). Similarly, determination of irreducible water saturation was more accurate using T2 approach (R2 = 0.87), whereas features extraction technique also exhibited a decent performance (R2 = 0.71). Also, T2 approach is more convenient since it is more straightforward to generate T2 spectrum from NMR measurements and use it for the ML models. Furthermore, based on the machine learning approach, gradient boosting and deep neural networks models performed with higher accuracy than other algorithms. This can be attributed to their strong configuration, which is able to find patterns between input and output parameters. This study provides more insight into petrophysical properties determined from NMR measurements in carbonates using ML techniques. This is useful in better characterizing carbonate reservoirs in the Middle East through accurate estimations of hydrocarbon resources and related reserves.