A Machine Learning Approach for Stress Prediction in Granitoid Formation at FORGE Geothermal Site Using Compressional and Shear-wave Slowness

The accurate estimation of in-situ stresses is of vital importance for optimizing the subsurface planning and design such as horizontal well placement, hydraulic fracturing, and wellbore stability. In this study, we apply machine learning models for interpreting well logs for stress prediction based on training with laboratory generated triaxial ultrasonic velocity (TUV) data. A total of 46 TUV data points were utilized that contain P and S-wave slowness (inverse of velocity) in x, y, and z-directions under different combinations of applied stresses. The experimental dataset represents the Granitoid formation retrieved from well 16A(78)-32 of the Utah FORGE geothermal site. Machine learning prediction models were developed for estimating stress using three broadly accepted techniques such as functional network (FN), adaptive neuro-fuzzy inference system (ANFIS), and artificial neural networks (ANN). The prediction results of all three models revealed high accuracy and reliability in terms of high coefficient of correlation (R) and low root mean squared (RMSE) errors. A comparison of prediction performance exhibited that ANFIS model outperformed the FN and ANN models demonstrating highest R value of 0.985, 0.972, and 0.976, and lowest RMSE of 2.34, 2.7, 3.4, for the testing and validation of predictive models of vertical and two horizontal stresses. Further, the optimized ANN model is transformed into a mathematical model for stress estimation using sonic log velocity data.