An Integrated Machine Learning Workflow to Estimate In Situ Stresses Based on Downhole Sonic Logs and Laboratory Triaxial Ultrasonic Velocity Data

Abstract

The optimum performance of various subsurface operations such as stimulation treatments, wellbore drilling, horizontal well placement, underground mining, and tunneling rely on accurate estimation of the in situ stresses. This study presents an integrated machine learning (ML) workflow for the reliable determination of the in situ stress in subsurface rock formations. The study workflow was completed in three phases. In the first phase, six supervised ML regression techniques were employed to develop the stress prediction models using laboratory true triaxial ultrasonic velocity tests (labTUV) data of subsurface rock samples retrieved from well 16A(78)-32 located at the Utah FORGE geothermal site. In the second phase, subsurface geological formations were classified into rock facies using an unsupervised K-means clustering algorithm. Finally, in the third phase, the optimized ML models were employed for predicting the in situ stresses in the corresponding rock facies in the well 16A(78)-32 using field sonic logs. A comparison of evaluation metrics revealed the superior performance of ANN models for horizontal and vertical stress predictions with root mean squared error of 1.5, 0.6, and 1.7 MPa, and determination coefficient (R2) of 0.98, 0.98, and 0.96, for the testing/validation phases, respectively. The generalization capability of ML models was explored by uncovering the underlying physics. The mathematical expressions of constitutive relations were extracted from three ANN models. Further, a total of five rock facies were identified in the subsurface geological formations. The novel workflow would be capable of delivering reliable in situ stress profiles in subsurface geological formations without performing expensive field tests.