High-resolution seismic impedance inversion integrating the closed-loop convolutional neural network and geostatistics: an application to the thin interbedded reservoir

Abstract

Abstract Seismic impedance inversion is one of the key techniques for quantitative seismic interpretation. Most conventional post-stack seismic impedance inversion approaches are based on the linear theory, whereas the relationship between seismic response and impedance is highly nonlinear. Thus, it is challenging to implement conventional inversion methods to obtain high-resolution impedance for reservoir investigation. Convolutional neural network (CNN), a superior deep neural network, has a strong learning ability, which can learn from data and establish complex nonlinear mapping. However, CNN-based methods are generally heavily dependent on amounts of labeled data. Hence, an alternative seismic inversion approach is proposed that combines the closed-loop CNN and geostatistics. The closed-loop CNN is less dependent on labeled data, characterized by utilizing labeled data and unlabeled data simultaneously to train the neural network. The two subnets represent forward modeling and inversion respectively, constraining each other during the neural network training. Geostatistics can be used to enrich the training data for neural network training, taking into account geological and geophysical prior information. Synthetic data testing reveals that the proposed inversion scheme can obtain more reasonable results benefiting from labeled training data augmentation. The proposed inversion scheme is applied to the field data for identifying thin interbedded reservoir within delta depositional system. The predicted results obtained by the proposed inversion scheme are consistent with well log data and geological settings, offering insights into reservoir characterization and hydrocarbon identification.