Machine Learning-based Prediction of In-Situ Stresses at the Utah FORGE Geothermal site

ABSTRACT: Information about the in-situ stresses plays a vital role in optimizing various subsurface operations including hydraulic fracturing, horizontal well placement, and stability of wellbores. This study presents a machine learning-based framework for predicting in-situ stresses in subsurface geological formations using laboratory-based ultrasonic wave velocities and field-based sonic logs. First, the relationship between ultrasonic wave velocities and stresses was explored by true triaxial ultrasonic velocity (TUV) test in the laboratory. TUV tests were performed on subsurface cores obtained from a geothermal well 16B(78)-32 drilled at Utah FORGE site. The TUV dataset was used to develop prediction models for three mutually orthogonal stresses using three robust machine learning (ML) algorithms including multi-layer perceptron (MLP), random forest (RF), and extreme gradient boosting (XGB). Further, an unsupervised K-means algorithm was employed to identify rock facies based on log-based rock attributes in the same well. Subsequently, ML models with the best prediction performance were employed to estimate in-situ stresses in the representative rock facies of subsurface rocks using field sonic logs of the same well. Although all three predictive models demonstrated reliable and consistent performance, accuracy measures revealed the superior performance for the MLP models with the determination coefficient (R2) of 0.98, 0.973, and 0.977, and root mean squared error (RMSE) of 2.88, 2.01, and 1.89 for the testing phases of the vertical, minimum horizontal and maximum horizontal stress models, respectively. A comparison showed that ML models predicted in-situ stresses in agreement with geomechanical model-based in-situ stresses. The proposed ML-based approach is capable of capturing the trends and variability in the stress and acoustic velocities and thus, provide reliable and cost-effective solution of estimating in-situ stresses in the subsurface rock formations without executing costly field-based injection tests.