Time-lapse seismic monitoring of elastic reservoir parameters and saturation using multitask and transfer learning

Time-lapse real-time monitoring of reservoir changes is essential during CO2 injection or hydrocarbon production. Here, we leverage convolutional neural networks (CNNs) that employ multitask (MTL) and transfer (TL) learning with the Gaussian Error Linear Unit (GELU) activation function to accurately predict relevant reservoir parameters from time-lapse seismic data. CNNs are initially trained to estimate the P-wave velocity from 2D multicomponent seismic data and then are fine-tuned through TL to obtain the VP/VS ratio, acoustic impedance, and saturation. This methodology is applied to a synthetic CO2 sequestration model based on California’s Kimberlina storage reservoir. When using MTL, CNNs are trained simultaneously on several related tasks by taking advantage of their commonalities. We show that after pretraining the model on a 2D line, it can be fine-tuned to predict the reservoir parameters from the data acquired in the crossline direction. Our work addresses the challenge of training-data scarcity by first training the network on a physics-informed synthetic data set and then fine-tuning it via transfer learning on the available data. The key contribution is a data and compute-efficient workflow, in which a pretrained P-wave model is reused through transfer learning to recover several elastic properties and saturation from limited training data. Adopting the GELU activation improves the generalization of the pretrained model and reduces the number of epochs. The predictions remain accurate for data contaminated by realistic noise. Our methodology promotes efficient use of computational resources in reservoir monitoring and helps increase the accuracy of real-time monitoring of the fluid movement inside the reservoir.