Seismic elastic parameter inversion via multi-attribute assisted multi-trace supervised learning

Abstract Elastic parameters (primary wave velocity V P , shear wave velocity V S , and density) are crucial to link seismic data and the corresponding petrophysical properties. Traditional elastic parameter inversion methods require high quality seismic data and face challenges in the suitability of physical models, the decoupling of coupled parameters and high-resolution characterization. Deep learning method, as a nonlinear powerful tool, gains more attention for precise elastic parameter inversion. To improve the generalizability of the supervised optimization model, we propose multiple attributes (partial angle stacks, partial angle stack differences and elastic impedance) assisted elastic parameter inversion method. The input data utilize amplitude variation characteristics from both seismic data and elastic impedance. In terms of network training, we first design an overlapping patching strategy during data preparation to augment the labeled data. Then, a special Unet, which performs max-pooling only along the time axis and employs a strip-like Conv2d kernel, is used for multi-trace inversion, during which a masking strategy is introduced to cope with single-trace labels. Besides, we employ attribute filtering and modified tanh activation function strategies to improve the performance in accuracy and horizontal continuity. Both synthetic and field data tests demonstrate that the proposed multi-attribute assisted multi-trace elastic parameter inversion method exhibits superior accuracy, robust noise resistance, and satisfactory horizontal continuity.