ESTIMATING IN-SITU STRESS MAGNITUDES FROM SONIC LOG DATA USING A MACHINE LEARNING MODEL TRAINED TO LABORATORY TRIAXIAL ULTRASONIC VELOCITY EXPERIMENTS

The reliable estimation of in-situ state of stress in the sub-terranean rock formations is crucial to address various field challenges such as hydraulic fracturing, wellbore stability, reservoir depletion, and induced seismicity. Here we present a machine learning (ML) predictive model for vertical and horizontal stress magnitudes in with a case study from the Utah FORGE geothermal site. Field data is obtained from sonic logging and includes compressional (P-) wave, fast shear (S-) wave, and slow shear wave slownesses (or their reciprocal values, the velocities). The ML models are used to estimate the in-situ stresses owing to the dependence of the wave velocities (i.e. 1/slownesses) on stress state. The ML models were trained on a laboratory experimental dataset quantifying true triaxial ultrasonic velocities (TUV) on rock sampled from the same interval that is to be evaluated using the ML model. Predictive models for vertical and horizontal stresses were developed using broadly accepted machine learning tools including support vector machine (SVM), decision tree (DT), random forest (RF), extreme gradient boosting (XGB) and adaptive gradient boosting (AGB). All the ML models of stresses exhibited good prediction performances in terms of high coefficient of determination ‘R²’ and low average absolute percentage ‘AAPE’ and root means squared errors ‘RMSE’. The evaluation metrics revealed that XGB and RF prediction performances are relatively better than other ML techniques reflecting R² of 0.95 and 0.94 for the testing and validation dataset. The AAPE and RMSE of testing and validation of models were observed to be 3.4 and 4.7 for XGB and 3.9 and 5.3 for RF, respectively. Based on these promising results, the ML models appear capable of providing estimates of all in-situ stress magnitudes. These estimates are provided in a continuous manner that is able to track variability along the wellbore. Ongoing work is aimed at increasing the quantity of training data, further optimizing the ML model for better fidelity to the training data, and using clustering algorithms to differentiate zones where velocities change primarily due to changing rock type and therefore should not be interpreted as stress variability.