Apparent Velocity Extraction from Quadrupole Acoustic Logging While Drilling: CNN-Based Moveout Regression

Acoustic logging while drilling (LWD) provides formation shear-wave velocity in real time during drilling, supporting geosteering and geomechanical evaluation. Shear-wave velocity is currently extracted by slowness–time coherence (STC) processing, which relies on a grid search for the global extremum; its hard decision is prone to dense outlier spikes in multi-event wavefields, and its computational burden is unfavorable for downhole real-time evaluation. We propose a lightweight convolutional neural network (CNN) that regresses the inter-receiver moveout directly from array waveforms; five-fold cross-validation on a synthetic dataset yields a median relative velocity error of 2.51% ± 0.16%. The numeric and image-inversion domains, with either a threshold mask or an Attention U-Net segmentation mask, give consistent velocities, with differences between mask methods of ≤0.6%, and 2.5–4.5 kHz band-pass filtering brings the two domains into convergence. As a control, supervised regression labeled by a measured STC log attains per-gather r = 0.72 over the full well interval under random splitting, but r drops to 0.20 under depth-blocked cross-validation that removes near-neighbor leakage, delimiting the applicability boundary of this proxy route. Inference is a single deterministic forward pass, the model occupies only 266 KB, and it runs approximately 25× faster than vectorized STC, making it suitable for edge deployment.