Method and application of sand body thickness prediction based on virtual sample-machine learning

Abstract

The purpose of this paper is to clarify the spatial spreading characteristics of the channel sand body in the Jurassic Shaximiao Formation reservoir in central Sichuan, and to improve the precision of channel characterization. Aiming at the problems of insufficient machine learning training samples and a lack of continuity of prediction results in the study area, we select the No. 7 sand formation of the second member of Shaximiao Formation as an example and use the method of combining boosted regression tree (BRT) model and virtual points to accurately depict the spatial distribution of the sand body. Starting from the known sand thickness and seismic attribute data, the BRT model is used to fuse the selected attributes to obtain the preliminary prediction results. On this basis, grid division is used to select virtual points to obtain three virtual datasets for sand body prediction. The three predictions are then analyzed using the clustering?topology method to obtain the dominant regions, and the virtual points are selected a second time for the final sand body prediction. The results show that the prediction accuracy of the BRT model is improved compared with other machine learning methods. Meanwhile, to address the insufficient number of samples in the study area, after using the two-stage virtual point generation method proposed in this paper, the R² of the test set in the model training results reaches 0.887. The final prediction results show that the sand body distribution effect is satisfactory, the lack of continuity of the channel can be improved, and the agreement with the well is high.