Deep learning for predicting water saturation using rock physics analysis and geostatistics theory: A case study of the P_sh⁸ in GFZ area, Ordos Basin

Tight sandstone reservoirs have strong heterogeneity and complex gas–water relationship, causing difficulty in quantitatively predicting water saturation. Deep learning, combined with rock physics analysis and geostatistics theory, was used to predict water saturation in tight sandstone, focusing on the P_sh⁸ in the GFZ area of the Ordos Basin. Results show that: (1) Starting with actual wells where porosity and saturation results are obtained from log interpretations, the relationship between reservoir parameters (porosity and saturation) and elastic properties (P-wave velocity, S-wave velocity, and density) is established through the development of a rock physics model suitable for the region. Under the constraints of geostatistical laws, such as background trends of elastic and reservoir parameters and the vertical variations in logging curves, reservoir conditions (including porosity, saturation, and thickness) are simulated to generate numerous pseudowells and corresponding seismic gathers modeled using the Zoeppritz equation. A convolution neural network is used to train the target curve and predict the target body. The predicted water saturation of the Psh8 shows strong agreement with the results from two blind wells, providing a reliable basis for understanding the water saturation (Sw) of tight sandstone.