Direct Inversion of Wellbore Acoustic Waveforms Using Convolutional Neural Networks

Waveforms acquired from acoustic tools in wellbores are essential for determining formation elastic and geomechanical parameters that are required for subsurface energy production applications such as carbon storage and geothermal energy. This data is also used for well monitoring and to verify the integrity of reservoirs throughout the lifetime of a wellbore. Traditionally, waveforms are measured by a set of receivers on a logging acoustic instrument and are processed using time- or frequency-domain methods to yield the formation acoustic modes, from which a subset of formation properties are inferred using inverse modeling. However, noise interference can affect the accuracy of these methods, and the processing typically requires significant computational resources and time. This paper presents a novel machine learning-based methodology for directly inferring formation properties from raw waveform data, eliminating the need for traditional acoustic modeling. The approach is applied to isotropic and vertical transversely isotropic (VTI) formations using data from dipole sources on wireline acoustic instruments. Results indicate that the methodology is most effective when applied to time- domain data using convolutional operations. Among the tested machine learning models tested, the convolutional neural network (CNN) with batch normalization demonstrated the highest efficiency and accuracy. The proposed method yields predictions with a mean absolute relative error of 2% for shear and compressional velocities in isotropic formations and below 4% for each of the five velocities in VTI formations. The proposed methodology enables direct prediction of formation properties from raw waveforms and can be applied to any acoustic source or instrument, whether for use during drilling or for post-drilling wellbore monitoring.