A combined approach for estimating horizontal principal stress magnitudes from borehole breakout data via artificial neural network and rock failure criterion. (December 2020)

• Record Type: Journal Article
• Authors: Lin, H.; Singh, S.; Oh, J.; Canbulat, I.; Kang, W.H.; Hebblewhite, B.; Stacey, T.R.
• Is Part Of: International journal of rock mechanics and mining sciences. Volume 136(2020)
• Publication Date: 2020-12
• Subjects: Horizontal stress estimation -- Borehole breakout -- Artificial neural network -- Failure criterion -- True triaxial test; Rock mechanics -- Periodicals; Soil mechanics -- Periodicals; Mining engineering -- Periodicals; Roches, Mécanique des -- Périodiques; Sols, Mécanique des -- Périodiques; Technique minière -- Périodiques; 624.151305
• DOI: 10.1016/j.ijrmms.2020.104539
• Languages: English
• ISSNs: 1365-1609

Abstract

Abstract: In this paper, a newly proposed approach on horizontal stress estimation from borehole breakout data is presented. In a previous study, a machine learning model was developed, capable of estimating maximum horizontal stress ( σ H ) from breakout data accurately. However, due to the limitation in experimental data, it was difficult to obtain the minimum horizontal stress ( σ h ) reliably. In this study, a series of breakout tests on Hydrostone-TB specimens was carried out to investigate the influence of σ h and vertical stress ( σ v ) on breakout geometries, as these two parameters were rarely studied previously. Results revealed that both breakout angular span and depth decrease with increasing σ h or σ v, although the influence of σ h is more significant. Based on experimental results from this paper as well as the literature, nine failure criteria were examined on the prediction accuracy of σ h providing the magnitude of σ H . Except for one model, all the other eight failure criteria consider the influence of σ v, as indicated in experimental findings. However, none of the failure criteria yielded reasonable σ h estimations. To overcome this problem, an Artificial Neural Network (ANN) model was developed from the experimental dataset. Once the model was constructed, it was examined against twenty-three field data, and yielded an acceptable average error rate of 15.88% on σ h considering the easily accessible breakout data. Then a comparative analysis on σ H estimation was performed via a number of approaches, namely, Kriging, ANN, and constitutive modeling. Results revealed that the use of the Mogi-Coulomb failure criterion is the most reliable approach for σ H estimation, with an average error rate of 6.82%. Overall, this newly presented 'ANN'-'Mogi-Coulomb' approach to horizontal stress estimation shows reasonable prediction results, which is expected to be improved in future studies by including additional data.