Inverting Elastic Model Properties Using ResNet

Summary We develop a novel seismic data inversion method to estimate the properties of subsurface layered elastic models using Convolutional Neural Network (CNN). Specially, we use ResNet (Residual Neural Network) to predict the parameters of layered elastic models, including layer depth, layer density, P-wave, and S-wave velocities for its unique identity block architecture. The entire dataset consists of 10,000 layered elastic models and their corresponding single shot records. We use 80% pairs from the dataset to implement the training process, and then the trained network could make predictions on the rest of the models in the dataset. Our trained network presents satisfying prediction results on both simple (i.e., few-layer) and complex (i.e., multi-layer) models, and it suggests that the proposed approach could be a useful tool for data processing, especially when dealing with near-surface layered models.