Improved Reservoir Rock Porosity Prediction from Advanced Mud Gas Data

Abstract

Abstract In our continued effort to extend the utility of advanced mud gas (AMG) data from the traditional fluid typing to reservoir rock properties prediction, this study investigates the feasibility of predicting the full porosity log for the hydrocarbon-bearing zones of wells ahead of wireline logging and core analysis processes. Our previous incremental results have confirmed the successful prediction of missing porosity logs in an interval within the borehole and in a section of a field. We established the linear correlation between porosity and Total Gas (TG) to confirm the hypothesis. Leveraging the capability of machine learning (ML) algorithms to recognize hidden patterns in data, we developed artificial neural network (ANN), decision trees (DT), and random forest (RF) models. We collected over 20,000 data points from representative wells in the study area, used 90% for training and optimizing the models, 10% for testing, and five wells for blind validation. A cut-off of 500 ppm was applied on the total gas to remove background gas effects and focus on the hydrocarbon-bearing zones. Using statistical model performance evaluation metrics comprising correlation coefficient (R2) and mean squared error (MSE), we compared the results of the ANN, DT, and RF models. The RF model consistently outperformed the others based on the training, testing, and validation metrics. Using the original AMG data, the least-performing RF model gave an R2 value of 0.78 for training, 0.76 for validation, and MSE of 0.014 for full-well blind testing. After applying the cut-off, the performance of all the models improved significantly, while the RF model maintained its best performance. With this improvement, the least-performing RF model gave an R2 value of 0.98 for training, 0.89 for validation, and MSE of 0.003 for full-well blind testing. Considering the outcome of our previous studies, these results have further confirmed the robustness of nonlinear solutions based on the ML methodology. It can be concluded that the ML approach for predicting reservoir rock porosity from AMG data acquired in real time is feasible, though with room for improvement. This study also confirms the benefit of focusing on the productive zone by applying a cut-off on the TG. This study will contribute to the objectives of digital transformation in the petroleum exploration industry by (1) expanding the utility of existing data without extra cost, (2) utilizing real-time data such as the AMG to predict rock properties in real time for better decision, (3) providing more information to optimize reservoir contact while drilling, and (4) providing information to determine reservoir quality at the early stage of well development.