Flexural-dipole sonic logging has been widely used as the primary method to measure formation shear slowness because it can be applied in both fast and slow formations and can resolve azimuthal anisotropy. The flexuraldipole waveforms are dispersive borehole-guided waves that are sensitive to borehole geometry, mud and formation properties, and therefore the processing techniques need to honor the physical dispersive signatures to obtain an accurate estimation of shear slowness. Traditional processing techniques are based on either a modeldependent method, in which an isotropic model is used as a reference to compensate for the dispersion effect, or a model-independent method, which optimizes nonphysical parameters to fit a simplified model to the field dispersion data extracted in the slowness-frequency domain. Many methods often require inputs such as mud slowness, frequency bandpass filter, or an initial guess of formation shear. Consequently, these methods often fail to interpret the dispersion signature properly when those inputs are inaccurate or when the waveform quality is poor due to downhole logging noises. The users must manually tune the processing parameters and/or choose different methods as a workaround, which causes extra time and effort to obtain the result hence imposes a significant challenge for automating sonic shear interpretation.

We develop a physics-driven machine learning-based method for enhancing the interpretation of borehole sonic dipole data for both wireline logging and loggingwhile- drilling. Extensive synthetic databases (i.e., lookup tables) are generated from an anisotropic root-finding mode-search routine and used to train neural network models as accurate and efficient proxies. Those neural network proxies can be used for real-time sensitivity analysis and performing inversion to the measured sonic dipole dispersion data to estimate relevant model parameters with associated uncertainties. Alternatively, various machine learning methods can also be developed based on the generated training dataset and that can be used for inferring relevant model parameters with uncertainties from the field data directly. We introduce how these trained models can be used to enhance the labeling and extraction of different dispersion modes. We developed a new method that outperforms previous modeldependent and model-independent approaches because the new method introduces a mechanism to constrain the solution with physics that also has the capability to incorporate more complicated physical dispersion signatures.

This new method needs neither prior information such as mud slowness and formation shear slowness, nor any tuning parameter to be played by the user. It also paves a way to automatically identify different anisotropy mechanisms such as intrinsic, layering, stress, or fractures. This leads to significant progress toward automated sonic interpretation.

The algorithm and workflow have been tested on field data for challenging borehole and geological conditions and compared with traditional flexural-dipole processing techniques with great success.

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