Machine Learning Regressors and their Metrics to Predict Synthetic Sonic and Brittle Zones

Gupta, Ishank (University of Oklahoma) | Devegowda, Deepak (University of Oklahoma) | Jayaram, Vikram (Pioneer Natural Resources) | Rai, Chandra (University of Oklahoma) | Sondergeld, Carl (University of Oklahoma)



Planning and optimizing completion design for hydraulic fracturing require a quantifiable understanding of the spatial distribution of the brittleness of the rock and other geomechanical properties. Eventually, the goal is to maximize the SRV (Stimulated Reservoir Volume) with minimal cost overhead. The compressional and shear velocities (Vp and Vs respectively) can be used to calculate Young’s modulus, Poisson’s ratio and other mechanical properties. In the field, sonic logs are not commonly acquired and operators often resort to regression to predict synthetic sonic logs. We have compared several machine learning regression techniques for their predictive ability to generate synthetic sonic (Vp and Vs) and a brittleness indicator, namely hardness, using the laboratory core data. We used techniques like multi-linear regression, lasso regression, support vector regression, random forest, gradient boosting and alternating conditional expectation. We found that the commonly used multi-linear regression is sub-optimal with less-than-satisfactory predictive accuracies. Other techniques particularly random forest and gradient boosting have greater predictive capabilities based on several error metrics such as R2 (Correlation Coefficient) and RMSE (Root Mean Square Error). We also used Gaussian process simulation for uncertainty quantification as it provides uncertainty estimates on the predicted values for a wide range of inputs. Random Forest and Extreme Gradient Boosting techniques also gave low uncertainties in prediction.