Prediction of nuclear magnetic resonance porosity well-logs in a carbonate reservoir using supervised machine learning models

Abstract

Porosity estimation is a fundamental input for reservoir management and petrophysical characterization, and this feature is usually estimated based on laboratory measurements or through the use of well-logs. As an important resource for porosity quantification, nuclear magnetic resonance (NMR) well-logs are extremely useful; they allow geologists and petrophysicists to rapidly quantify different types of porosities (including total, effective, and free fluid porosity), and to perform a full formation evaluation and a reservoir quality analysis. However, the activation of wireline tools, the signal-to-noise ratio, the environmental conditions, and the characteristics of the formation fluid can create expensive and adverse conditions for subsurface acquisition. This research aims to develop machine learning models for the creation of synthetic NMR well-logs, assisted by auxiliary well-logging features. Four supervised models: multilayer perceptron neural network, AdaBoost, XGBoost, and CatBoost, comparing the adjusted R2 and RMSE. Of these, the CatBoost regressor provided the most highly optimized model. It was able to reduce local dissimilarities with the real dataset, and returned a better global metric score, yielding an adjusted R2 of 0.87 and an RMSE of less than 0.01. Moreover, all of the machine learning models provided substantial improvements in total porosity estimation, particularly compared to conventional empirical calculations based on density and sonic well-logs. An improvement of 0.5520 in the adjusted R2 was achieved for the density porosity, and 0.2 for the sonic porosity. The differences between real NMR well-logs and the machine learning outputs were in general less than 5%, for most of the well-logging interval. In addition, a tree boosted porosity model based on well-logs is presented for the first time, and the contributions and impacts of the input features on the model predictions are explored. Finally, the behaviors of the linear and nonlinear features of the model are examined, which allows us to better understand the complex relationships among the features and the dataset used.