Adaptive Gated Compensation Network for Shear Wave Velocity Prediction

Abstract Accurate shear wave velocity (VS) estimates are essential for seismic inversion and reservoir characterisation. The cost and operational requirements of dipole sonic logging can limit the availability of measured VS. Predicting VS from conventional well logs is a practical alternative. Empirical relationships can be sensitive to geological setting, whereas deep learning models trained on limited data may predict less accurately in wells excluded from training. This paper presents an Adaptive Gated Compensation Network (AGC-Net) that combines a physics-inspired linear shortcut with a gated deep sequence pathway. The linear branch provides a baseline estimate, while temporal convolutions and attention model nonlinear corrections. An input-dependent enhancement gate scales these corrections, and a composite loss combines mean squared error with a concordance correlation coefficient (CCC) loss. On the blind test well, AGC-Net outperformed eight baselines, achieving a coefficient of determination (R2) of 0.808. The root mean square error and mean absolute error were 70.5 m/s and 53.9 m/s, respectively. Ablation experiments supported the contributions of the linear shortcut, gated enhancement and CCC loss within the evaluated split. Removing the linear shortcut caused the largest decline in blind-well accuracy.