Reconstruction of missing shear wave slowness in shale gas reservoir based on deep learning algorithm

Abstract

Shear wave slowness (DTS) is a significant data for reservoir evaluation and description. Due to the cost and time-consuming, the missing of DTS is inevitable in well logging. Empirical equations and traditional machine learning methods are often used to estimate the general trend of DTS, but their ability to predict the local details is limited, which is particularly important for highly heterogeneous reservoirs such as shale. In this paper, based on the generative adversarial network (GAN), a new deep learning model was established for DTS reconstruction, in which the long short-term memory network (LSTM) was applied as the generator to learn the overall trend of logging data and generate pseudo DTS sequence, the one-dimensional convolutional neural network (1DCNN) was applied as the discriminator to capture the local details of logging data and judge the authenticity of DTS sequence. The generator and discriminator in GAN were trained at the same time to achieve model convergence and the high distribution similarity between reconstructed DTS and the real one. Several data experiments were conducted with the logging data of 102 wells from Weiyuan shale gas field, China, and the proposed model were proved to have good prediction ability for both the overall fluctuation and local details of DTS logging data. The processing flow recommended in this paper provides an economic and fast choice for the reconstruction of missing DTS in shale gas reservoir and shows the potential of applying deep learning algorithm in assisting logging data generation.