Prediction of NMR T2 Macro- and Micro-Porosity with Machine Learning Techniques: Considering the constraints of ECS Lithology Classification

Abstract

Abstract Nuclear magnetic resonance (NMR) scanning, particularly real-time NMR Logging While Drilling (LWD), offers a non-radioactive approach for porosity measurements. As a primary technology for pilot well logging, NMR loggings has a high cost, so it becomes imperative to develop alternative cheap and efficient models to predict NMR-derived porosity using conventional well logs. With the advances in computational power, Machine Learning (ML) has become promising to tackle a wide range of complex engineering and scientific problems while striking a good balance between accuracy and efficiency. This work aims to develop a machine learning-based workflow to predict T2 macro-porosity and micro-porosity without expensive NMR logging information. We propose to enhance the accuracy of the prediction by considering the rock-typing classification obtained from Elemental Capture Spectroscopy (ECS) logging. We collect 25534 data samples within a depth interval of 2900 feet in a mixed siliciclastic-carbonate reservoir. Through ECS lithology interpretation, we identify four distinct rock types, including organic-rich shale, non-organic calcareous shale, calcareous siliciclastic, and shaly carbonate. We evaluate the distributions, importance rankings, and correlation coefficients for the potential input variables for the ML models and identify the critical input features, including gamma ray (GR), neutron porosity (NPHI), bulk density (RHOB), deep lateral resistivity (LLD), compressional wave slowness (DTC), and photoelectric factor (PE). We then separately train a variety of ML models for each lithofacies category to enhance prediction accuracy. For comparison, we also implement ML models without considering lithofacies constraints. We examine the performance of the ML models using various accuracy tests, including predictive cross-plots, coefficient of determination (R2), and mean square error (MSE) methods. Our findings indicate that adaptive gradient models outperform other ML techniques, such as random forest, extreme gradient boosting, and nearest neighbor models. Besides, after introducing the lithology interpretation into the ML models, the R2 score for predicting T2 micro-porosity significantly improves, jumping from a mere 0.192 to a robust 0.952. Similarly, the R2 score for predicting T2 macro-porosity increases substantially, climbing from 0.653 to an impressive 0.967. This underscores the crucial role of factoring in lithology classification for petrophysicists when leveraging conventional well log data for porosity predictions, especially for complex lithology reservoir. This study establishes an ML prediction workflow for NMR T2 macro-porosity and micro-porosity while considering the constraints of ECS-based lithology classification. Moreover, For the T2 macro- and micro- porosity of the four lithologies, the MSEs of adaptive gradient model are less than 0.2. It provides a rapid and accurate tool for estimating rock porosity at a granular level, thereby guiding subsequent engineering decisions, including those related to drilling and completion processes.