Deep-Learning-Based Prestack Seismic Inversion Constrained by AVO Attributes

Prestack seismic inversion is an effective approach to obtain elastic parameters for reservoir characterization in seismic exploration. However, the difficulty in achieving reliable inversion results in prestack seismic inversion remains due to the strong nonlinearity of the problem and the ambiguity of solutions. Deep learning (DL) excels at mapping complex nonlinear relationships, and thus various DL-based methods have been used in seismic inversion. To address challenges posed by the nonlinearity and ambiguity in seismic inversion, a novel DL-based prestack seismic inversion constrained by amplitude versus offset (AVO) attributes is proposed. In this approach, the multitask learning (MTL) strategy is adopted to construct a deep neural network that allows for the simultaneous processing of multiple related tasks through information shared between tasks. Moreover, the theoretical seismic forward modeling is integrated with the neural network training, enabling semisupervised learning and utilizing unlabeled data. Additionally, to mitigate the ambiguity of solutions, AVO attributes including intercept
$P$
and gradient
$G$
are introduced as constraints in the neural network training process. Experimental analyses show that the proposed method can obtain superior inversion results on both synthetic and real examples. Compared with other DL-based methods, the mean squared error of the proposed method’s inversion results on examples drops by at least 30%. Besides, the proposed method can effectively improve spatial continuity and preserve more details laterally in the field data example.