Deep-Learning-Based Seismic Variable-Size Velocity Model Building

Abstract

Current data-driven inversion methods based on deep learning (DL) use an end-to-end learning to obtain a mapping relationship from seismic data to the velocity model. This method requires truncating the feature maps in the output layer of the network to build the velocity model with a specified size. Therefore, it lacks the flexibility of handling output velocity models of different sizes, as the network needs to be retrained to achieve reliable results when changing the desired size of the velocity model. Here, to solve the problem of data-driven velocity inversion with variable size output, a novel sampling matrix is proposed to compress the seismic data to the same size as the model to avoid clipping the feature maps. The compressed seismic data is fed into the designed deep convolutional neural network (CNN) model for training. Here, the network has a multi-scale encoder-decoder structure and is composed of several residual blocks. Also, the multi-objective loss functions balance strategy is used to optimize the training process. Utilizing the trained network to test compressed synthetic seismic data of various sizes, the effectiveness of the proposed method for inversion of variable-size elastic wave velocity models is verified.