Sparry calcite identification in shale reservoirs by data augmentation and attention-based machine-learning algorithms

Abstract Sparry calcite is an indicator of favorable production lithology, and it also exerts a positive influence on production capacity. Therefore, sparry calcite identification is an important part of shale reservoir exploration. However, the logging response to sparry calcite is weak and sample sizes are limited, making accurate identification challenging using conventional methods. To address this challenge, we develop data augmentation and attention-based machine-learning algorithms, specify logging curves as experimental data. Initially, the data set is optimized using synthetic minority oversampling techniques and wavelet transforms. Subsequently, an attention-based convolutional neural network is trained to extract nonlinear features from the input logging curves. These nonlinear features are then fed into the bidirectional long short-term memory to extract potential information regarding the depth domain sequence and achieve accurate identification of sparry calcite. Our method has been validated using a data set from shale reservoirs in the Dongying Depression, achieving an accuracy of 96% and exceeding 90% for recall and precision metrics. Furthermore, an adjacent-well experiment accuracy can reach 91%, precision 90%, and recall 85%. This model effectively integrates intelligent algorithms, resulting in significant enhancements in sample optimization and feature extraction.