Unsupervised-learning based self-organizing neural network using multi-component seismic data: Application to Xujiahe tight-sand gas reservoir in China

Abstract

The tight sandstone reservoirs of the Xujiahe Formation of the Western Sichuan Basin Depression typically exhibit characteristics such as low porosity, low permeability, and strong heterogeneity, resulting in weak seismic response differences between gas and water. In present study, we developed an unsupervised learning gas reservoir prediction method using a self-organizing neural network (SOM) to identify and predict tight-sand gas reservoirs in areas with few or no wells. First, we extracted seismic attributes data and obtained multi-component seismic composite attributes. Then, the multi-component composite attributes were fed into the SOM network for training, obtaining the gas reservoir prediction results. Simultaneously, the designed SOM prediction scheme was also applied to PP-wave and PS-wave seismic data respectively for gas reservoir identification and prediction. The prediction results of the multi-component composite attributes exhibited good correspondence with available gas drilling information, indicating their unique merits in resolving issues pertaining to the deep tight gas sandstone reservoir identification. Finally, based on geological data and drilling information, we evaluated the effectiveness of the prediction results of proposed methods, and predicted favorable exploration areas. Effective implementation contributes substantially to tight-sand gas reservoir identification and prediction in less or even no well area. • Unsupervised learning-based self-organizing neural network. • Delineate a tight-sand gas reservoir using multi-component seismic data. • Gas reservoir identification and prediction in few wells or even no well area.