Known for its highly anisotropic character, the Tuscaloosa Marine Shale (TMS) is an economically challenged formation in need of improved stress estimates. To avoid steep declines in production, hydraulic fracturing is used to release the hydrocarbons trapped in the matrix of the shale. Hydraulic fracture propagation and the overall hydraulic fracturing process depend on an accurate estimation of Young's modulus and Poisson's ratio. Moreover, to predict stress in horizontal laminated shales, several vertical transverse isotropic (VTI) models were proposed since ANNIE. However, such analytical methods require extensive calculations and knowledge of dynamic-to-static ratios when estimating mechanical properties from well logs. Machine learning (ML) can be applied to generate geomechanical synthetic logs and therefore, may represent an alternative to the current anisotropic techniques. In this paper, the Gradient Boosting algorithm is used to estimate the static properties of the Tuscaloosa Marine Shale.
The ML model is trained on a dataset containing conventional and geomechanical logs from two wells. The latter are obtained using an extension of the ANNIE method which handles the difference between the vertical and horizontal Poisson's ratios by employing empirical correlations from ultrasonic core data for the stiffness coefficients C13 and C11. To assure training data reliability, dynamic-to-static conversion ratios are applied and triaxial experimental data are used to validate the results. To predict the static mechanical properties of the shale formation, the proposed ML model uses the depth to account for compaction and three additional predictor features represented by the compressional slowness, and vertical and horizontal shear slowness logs.
The model successfully predicts the vertical and horizontal Young's moduli and Poisson's ratios, as well as the minimum horizontal stress. The coefficient of determination (R2) and normalized root mean square error (RMSE) are used to assess the performance of the model on the test dataset. Best R2 (0.99) and normalized RMSE (0.015) are obtained for the horizontal Young's modulus and minimum horizontal stress, respectively, while lowest R2 (0.96) and highest normalized RMSE (0.035) are computed for the vertical Poisson's ratio and vertical Young's modulus, respectively. Moreover, the model is used to estimate the static properties of a third TMS well without geomechanical logs. The predicted well logs show a good match to the available triaxial core data points, proving the applicability of the proposed model to the evaluation of the Tuscaloosa Marine Shale.
Machine learning eliminates all tedious steps associated with the calculation of stiffness coefficients and the need to convert dynamic properties to static. The presented ML model can be run in most well log analysis platforms by anyone who disposes of a complete suite of sonic logs and seeks to better understand the mechanical properties of other TMS wells.
This work creates a bridge between analytical methods and machine learning. Both petrophysical and geomechanical input was necessary, while results can further aid completion engineers in selecting the best intervals for hydraulic fracturing.