Real-Time Hydrocarbon Volume Fraction Prediction Using Machine Learning: A Sensitivity Analysis of Different Data Sources

Hydrocarbon volume fraction (HVF), the ratio of the hydrocarbon content of a rock to the total volume of all fluids and usually expressed as a fraction, is one of the essential reservoir properties, similar to water saturation, used for the reservoir volumetric assessment. The traditional methods of estimating HVF from core analysis and wireline log analysis have limitations. This study, through a robust sensitivity analysis, investigates the data source that best predicts HVF. We investigated the individual predictive capability of drilling parameters, advanced mud gas measurements, and cuttings lithology to predict HVF by running three groups of experiments. Datasets from five wells were used to train three machine learning (ML) algorithms: artificial neural networks (ANN), decision tree (DT), and random forest (RF). The database is randomly split into training and testing subsets in the 70:30 ratio. We also implemented the multivariate linear regression (MLR) algorithm as a benchmark. Each group were tested for their respective capability to predict HVF. The comparative results showed that the drilling parameters and mud gas data are capable of predicting HVF with the RF model emerging the best in both cases. The RF model recorded the highest training and testing R-Square values of 0.98 and 0.96 for the drilling parameters and 0.99 and 0.94 respectively for the mud gas data. The use of cuttings lithology to predict HVF was found to be unpromising with the ANN model recording the highest training and testing R-Square values of 0.30 and 0.27 respectively. Within the drilling parameters group, the MLR model gave the least training and testing R-Square values of 0.47 and 0.38 respectively in favor of the ML models with much higher values. This indicates that the relationship between the drilling parameters and HVF is nonlinear. Hence, HVF could not be accurately predicted using a linear model such as MLR. In the mud gas data group, all the models performed well, but the ML models did much better than the MLR. For the cuttings lithology group, all models performed poorly. This indicates that the ML methodology, along with other data-driven approaches, is still guided by the physical laws that are represented in the data. This study has proven that using either drilling parameters or mud gas data only will give the same level of accuracy, or even better, in real-time HVF predicting, among other benefits. This partly addresses the challenge of data availability and confirms that "less could be better". Our future work will investigate an optimal subset of features from all the real-time data that will produce the most accurate prediction of HVF.