Utilizing Advanced Mud Gas Data to Predict Oil Volume Fraction Using Machine Learning Algorithms

Abstract

Abstract Digital transformation is unleashing unprecedented potentials to maximize subsurface data utilization. Advanced mud gas data have been extensively used for reservoir fluid typing. Oil volume fraction (OVF) is traditionally obtained from core analysis and typically computed from wireline logs. Extensive core analysis is expensive and time-consuming. The underlying assumptions of the equations to compute OVF from wireline logs introduce uncertainties. This study proposes a machine learning (ML) approach to predict OVF based on advanced mud gas data to reduce uncertainties and increase accuracy. There is currently no established empirical correlation between advanced mud gas components and OVF, which creates the opportunity to apply ML algorithms. We applied two models comprising decision tree (DT) and random forest (RF). A multivariate linear regression (MLR) model was implemented as a benchmark. Advanced mud gas data and their corresponding OVF values from 10 wells were used. The gas data comprising 16 light and heavy components are the input features. All the wells were combined to build the dataset used to train the models, leaving one out at each run for blind validation. Out of the results obtained, we present two representative ones in this paper. We used correlation coefficient (R2) and mean squared error (MSE) as basic metrics to assess the performance of the models. The comparative analysis of the models’ performances showed that the RF model has the highest accuracy for both validation wells to predict OVF using advanced mud gas data. For well A, the RF model has a training R2 value of 0.98 and 0.79 for validation. For well B, the training and validation R2 values for the RF model are 1.0 and 0.84, respectively. The least performing model is the MLR, giving training and validation R2 values of 0.95 and 0.66 for well A and 0.95 and 0.67 for well B, respectively. This indicates that the challenge of predicting OVF using advanced mud gas data is more of nonlinear relationship that is better handled by a ML method rather than a linear model such as the MLR. The MSE for all the models are generally very low. The results also showed that the total normalized hydrocarbon is the most significant feature. The outcome of this study shows that predicting OVF from mud gas data is feasible using the ML methodology. ML methods are capable of handling the nonlinearity of the data even when the linear correlation is weak. There is also ample room for improving the accuracy of the models by adding more real time data. This study has showed the potential to predict OVF while drilling, ahead of core analysis and wireline data interpretation.