Cross-dipole sonic data have been widely accepted in the industry to characterize shear slowness measurements using borehole flexural modes. The borehole flexural mode is dispersive and sensitive to many factors such as tool structure, borehole shape, drilling fluid and formation properties. These sensitivities further enable us to characterize anisotropy, drilling features, and logging conditions. From wide-frequency-band dispersions, one can evaluate the cross-dipole dispersions to identify zones with intrinsic anisotropy, stress-induced anisotropy, or fracture-induced anisotropy. One can also check the dispersions to identify drilling features such as break-out and gas leak, and logging conditions such as tool decentralization. A good understanding of dispersion types is an essential step to guide petrophysical and geomechanical applications. Traditionally, the classification is often carried out manually through visual inspection on a depth-by-depth basis. This is mainly because the dispersion data are represented by scattered and noisy points. Consequently, such a process is time-consuming and requires knowledge on sonic dispersion signatures.

To automate sonic data classification, we have developed an algorithm using a supervised machine-learning approach. A neural network classifier was trained using a large volume of synthetic dispersions. The training data are generated using randomly sampled input physical parameters covering effective borehole ovality model, anisotropic model and an alteration model for near wellbore concentration effects. The training datasets are labeled with these physical models. Excellent accuracy was achieved with neural network training. We then applied the classifier to field prediction, where scattered dispersion points are first reconstructed as smooth curves using a recently developed automatic dispersion interpretation algorithm. We tested the new algorithm to real field data where we observe different formation types that are dominant. The testing results show that the developed workflow is accurate in predicting dispersion types with noisy data, and efficient in handling large volume wells.

The machine-learning based workflow provides a more natural way for classification than a criteria-based algorithm, thanks to the rich physical information brought by training datasets. The outputs of the classification give ‘soft flags’ in the sense they are represented by prediction probabilities between 0 to 1 using the concept of entropy, which further enable us to evaluate co-existing dispersion signatures. This workflow serves as a key step in the quality control of sonic data and guides further applications in geomechanics where features such as stresses, intrinsic anisotropy and breakouts are key inputs. This paper will describe the methodology, including the construction of the training library, and illustrate the classification with several field examples.

This content is only available via PDF.
You can access this article if you purchase or spend a download.