Yaju Hao

Nonstationary convolution models are widely used as the forward formula of nonstationary seismic reflectivity inversion (NSRI). This forward formula is determined by the convolution of a time-varying wavelet and reflectivity. Previous NSRI algorithms just focus on the estimation of reflectivity according to different sparse regularization strategies. The forward-operator, time-varying wavelet matrix is usually generated using a user-defined constant Q-value. However, the Q-value changes with time and space; therefore, inaccurate reflectivity inversion results will be yielded by the traditional NSRI method. In order to improve the accuracy of the inverted reflectivity, we propose a new forward-operator construction method based on the monotone relationship between the equivalent Q-value (Q_e) and centroid-frequency (CF). We first extract CF from the Gabor time-frequency amplitude spectrum of the seismic signal. Then, we can obtain Q_e through the CF-Q_e template. Hence, the blindness of selecting a Q-value is avoided and a more accurate forward operator can be constructed. Synthetic and field-data examples demonstrate that the proposed method provides a more accurate reflectivity estimation by constructing a more reliable forward operator. Furthermore, the enhanced reflectivity sections enable a clearer delineation of top and bottom interfaces of coalbed methane reservoirs, providing high-resolution seismic support for coal seam interpretation and subsequent mining planning.

Deep-water carbonate reservoirs are currently the focus of global oil and gas production activities. The characterization of strongly heterogeneous carbonate reservoirs, especially the prediction of fluids in deep-water presalt carbonate reservoirs, exposes difficulties in reservoir inversion due to their complex structures and weak seismic signals. Therefore, a multiparameter joint inversion method is proposed to comprehensively utilize the information of different seismic angle gathers and the simultaneous inversion of multiple seismic parameters. Compared with the commonly used simultaneous constrained sparse-pulse inversion method, the multiparameter joint inversion method can characterize thinner layers that are consistent with data and can obtain higher-resolution presalt reservoir results. Based on the results of multiparameter joint inversion, in this paper, we further integrate the long short-term memory network algorithm to predict the porosity of presalt reef reservoirs. Compared with a fully connected neural network based on the backpropagation algorithm, the porosity results are in better agreement with the new log porosity curves, with the average porosity of the four wells increasing from 89.48% to 97.76%. The results show that the method, which is based on deep learning and multiparameter joint inversion, can more accurately identify porosity and has good application prospects in the prediction of carbonate reservoirs with complex structures.