깊이별 분리 합성곱 신경망을 이용한 속도 모델 구축

Abstract

Accurate velocity model building is one of the most important tasks in seismic survey. Recently, as deep neural networks have gained huge popularity in the field of geophysics, studies have been published to predict velocity models using regular convolutional neural networks. In this paper, we proposed Tomography_CNN with depthwise separable convolutions encoder-decoder structure for velocity model building. This network is trained in supervised learning approach and we predicted P-wave velocity models not only from time-domain wavefields but also from laplace-domain wavefields. The depthwise separable convolutions, fundamental part of this network structure, perform spatial-oriented convolutions independently of each input channel. Depthwise separable convolutions can improve network performance as well as computational cost compared to regular convolutions. Synthetic velocity models generated for training contain a variety of geologic features, including stratigraphic structure, syncline and anticline structure, fault and salt-dome. We compared network trained with time-domain wavefields and network trained with laplace-domain wavefields using the same number of model parameters and same hyperparameters. Time-domain network had shown very promising results on test data and laplace-domain network had shown reasonable results on test data. In addition, time-domain network took a long time to train due to enormous data, but the prediction cost of one velocity model after training is negligible. Laplace-domain network could reduce the training time about 40% or more compared to time-domain network.