Few-Shot Learning Method for Continuous Prediction of Rock Mechanical Parameters Based on Logging Data.

Abstract

The stratigraphic sequence displays pronounced variations in lithological and physical characteristics along the vertical axis. The acquisition of vertically continuous variations in stratigraphic rock mechanical parameters is essential for the detailed characterization of stratigraphy, particularly in reservoir layers. Moreover, mechanical parameters are essential for the establishment of a rock mechanical stratigraphic framework. However, the existing research methods cannot obtain the complete stratum rock mechanics parameters by relying solely on a small number of rock samples. In this study, a method for continuously predicting rock mechanic parameters within geological sequences by using geophysical logging data was proposed. The method aims to solve the problems of sample scarcity and global geological feature extraction in the continuous prediction of rock mechanics parameters. This methodology achieves the continuous extraction of stratigraphic features, generation of samples, and prediction of the mechanical parameters. The research result showed that the stratigraphic feature extractor of the logging autoencoder possesses global stratigraphic perception capabilities. Stratigraphic feature extractors can precisely identify and pinpoint key locations where stratigraphic properties change, such as lithological transitions and abnormal parts within thick rock layers. The fake rock mechanical samples produced by the generator exhibit data characteristics identical to those of the genuine samples. The rock mechanics parameter prediction model with the ability to extract stratigraphic features performs significantly better than traditional models when evaluated on real data sets (the mean square errors for the friction angle and cohesion are 17.2 and 12.65, respectively, mean absolute errors are 2.85 and 2.64, respectively, and R² values are both 0.91) and produces more reasonable calculation results in the actual prediction task of Longtan Formation rock mechanics parameters compared to traditional models. This study provides an extensive applicable methodological framework to address the issues of insufficient samples and continuous prediction of geological parameters in the field of geology.