Tight Gas Sandstone Formation Lithology Classification Using Deep Learning Networks

Tight gas sandstone formation is an important reservoir type. Classification of rock facies based on borehole data is an important step in reservoir analysis. Considering the powerful learning ability of the neural network, many researchers have proposed various neural networks for lithology classification. An optimal network for lithology identification was selected by qualitatively comparing the effects of neural network categories, optimizer types, and the number of hidden layers on network prediction results. First of all, two commonly used neural network models are built for classifying lithology: back propagation neural networks (BPNN) and convolutional neural networks (CNN). Then, the well logging data from the tight gas sandstone formation is regarded as the model input to determine an optimal network for predicting lithology by comparing the confusion matrix and precision of the predicted results. The test in the tight sand reservoir demonstrates that compared with BP neural network, our neural network has better performance in lithology classification. In addition, to prove the rationality of our conclusion and the effect of CNN in other cases, the logging data of carbonate formation with more complex lithology are tested again, and the CNN model still performs well.