Gas-Bearing Prediction Using Transfer Learning and CNNs: An Application to a Deep Tight Dolomite Reservoir

Predicting gas-bearing zone of deep tight dolomite reservoirs from prestack seismic data is challenging and subject to great uncertainty. Machine learning especially for deep learning (DL) provides a new potential. One main limitation of the DL-based supervised methods is that they require large amounts of training data. However, well-log labels from the real deep reservoirs are very insufficient. To address this issue, we investigate a method based on convolutional neural networks (CNNs) considering transfer learning to predict gas distribution of deep tight dolomite reservoirs. The CNNs model we used contains three convolutional layers for automatic feature extraction from prestack data and one fully connected (FC) layer for gas-bearing probability prediction. A numerical model is designed based on petrophysical parameters extracted from the real target work area associated with deep tight dolomite reservoirs. The model is used to generate synthetic samples to pretrain the CNNs model. We then fix the network parameters in the first two convolutional layers and decay the learning rates of the third convolutional layer and the FC layer. Using the real samples to fine-tune the pretrained CNNs model with epoch increasing. The optimal predictor is finally trained well for gas-bearing prediction. The method is applied to a real work area of deep tight dolomite reservoir located in western China covering approximately 800 km
²
. Examples illustrate the roles of transfer learning on improving gas-bearing distribution of deep tight dolomite reservoirs and increasing the generalization of the method.