Building accurate acoustic subsurface velocity models is crucial for successful industrial exploration projects. Traditional inversion methods from field-recorded seismograms struggle in complex geological regions. Deep learning (DL) offers a promising alternative, but its robustness with field data in these regions needs further exploration. This study analyzes DL's ability to utilize labeled seismograms (field-recorded or synthetic) for accurate velocity model recovery in a challenging Gulf of Mexico region. The impact of training data selection and augmentation techniques on velocity profile recovery is evaluated. Models trained on field data outperformed synthetic data models based on Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM), and R-squared, yielding geologically plausible predictions and sharper migration images. Synthetic data models, while less precise, showed potential, especially with limited field data, but their effectiveness hinges on bridging the domain gap between training and test data using advanced wave equation solvers and geologic priors. The study underscores DL's potential for advancing velocity model-building workflows in industrial settings and highlights the role of earth scientists' expertise in curating synthetic data.

Stuart Farris, Robert Clapp, Mauricio Araya-Polo