Generative Adversarial Network-Based Framework for Accurate DTS Logging Curve Generation in Heterogeneous Reservoirs

Abstract Accurate generation of missing share wave slowness (DTS) logging curve is significant for the precise reservoir evaluation. While various data-driven prediction models have been proposed, only a few addresses the intricate details of the DTS curve shape, and it is significant for reservoirs with strong heterogeneity. In this study, a novel DTS generation framework consisting of generator and discriminator was established based on generative adversarial network. In the generator, with the input of compressional wave slowness and compensated neutron curves, the recurrent neural network was applied to gain insight into the general pattern and generate DTS curves. In the discriminator, the convolutional neural network was adopted to compare the detailed shape and evaluate the realness of generated DTS curves. Both the generator and discriminator underwent concurrent training, aiming for model convergence and achieving a close distribution resemblance between the generated DTS curves and authentic data. The proposed DTS generation framework was practically applied in a shale gas field in the Sichuan basin of China. By segmenting the complete logging curves from over 100 wells, 47200 sequences with a length of 32 were obtained in the dataset. After 50 rounds and 26900 training cycles, the generation model exhibited robust performance with an average relative error of 0.015, and a coefficient of determination of 0.91. The frequency distribution of the generated DTS value closely resembled that of the real ones, confirming the generation ability for both overall fluctuation and local detailed shape. Moreover, a blind test on logging curves in 8 wells revealed a high shape agreement between the generated and real DTS curves, indicating the applicability of the proposed generation framework. Unlike the conventional approaches emphasizing the overall trend of DTS curves, the proposed framework introduces an additional discriminator to enhance the generation ability for intricate local details, leading to significantly improved generation performance. This study underscores the potential of advanced artificial intelligence methodologies for precious logging curve generation.