Real-time formation fluid characterization is a key factor to an optimized hydrocarbon recovery strategy. Various techniques are available to address downhole fluid properties challenges. One of the major advances in formation evaluation technology is in the field of Mud-Gas Logging. Formation evaluation using mud gas analysis involves detection, quantification and analysis of the hydrocarbon liberated from the drilled rock into the drilling fluid. It provides the earliest insight into the hydrocarbon content and type in the formation. Technological advances have resulted in delivering the reservoir fluid composition of the lighter fraction from the mud gas measurements. Extraction efficiency computed using synthetic hydrocarbon mixtures prior to drilling help to deliver reliable composition of the hydrocarbon column in real-time. In this paper, a new digital workflow developed to interpret the mud gas data and to characterize reservoir fluid properties while drilling is discussed.
The first workflow is used to filter and remove background gas and unreliable data from further analysis. Most of the currently used analysis methods involve normalizing the hydrocarbon composition to the C1 to C5 range and using gas component ratios and cross plots for further analysis. To bridge the gap in the data analysis, we developed a second workflow using machine learning algorithms. This workflow uses a set of preconstructed machine learning models to estimate fluid type, C6+ composition and GOR. The models were trained using a proprietary PVT database that was compiled from laboratory fluid measurements. We selected samples by ensuring that the training data was reliable and representative of reservoir fluids from different geo-locations. Based on exploratory data analysis, we identified many features in mud gas composition which were used to build a random forest classification model for fluid type identification. We used Gaussian Process Regression to predict C6+ fraction from the mud gas data and subsequently the GOR of the reservoir fluids. The models were selected and trained to optimize the performance of the fluid prediction workflow in real-time application.
