An Invertible Neural Network-Based Elastic Parameter Seismic AVO Transfer Inversion Method

Abstract

Elastic parameters such as velocity and density play crucial roles in reservoir identification and hydrocarbon exploration. The seismic amplitude variation with offset (AVO) method is an important means of obtaining formation elastic parameters. Compared to model-driven AVO inversion methods based on convolutional forward modeling, data-driven methods grounded in neural network technology can provide elastic parameters with higher accuracy and resolution when sufficient training samples are available. However, in field work areas, well-log data suitable for creating training samples is often scarce, leading to insufficient training labels and thereby constraining the application of data-driven seismic AVO inversion methods. To address this issue, this paper proposes a seismic AVO transfer inversion method based on invertible neural networks for estimating P-wave velocity, S-wave velocity and density. Initially, an invertible neural network (INN) for inversion is pre-trained using a limited number of labeled training samples. Subsequently, the pre-trained network is migrated to the inter-well region based on unlabeled training samples. The proposed framework integrates the bijective mapping capability of the INN core with a two-stage transfer learning strategy, where a non-invertible low-frequency substitution layer with stop-gradient is used to inject geological priors. Tests on synthetic and field data demonstrate that the proposed method can effectively alleviate the dependency of data-driven seismic inversion on labels and achieves lower errors than the selected benchmark under the tested small-sample conditions.