Journal of the Korean Society of Mineral and Energy Resources EngineersISSN:2288-0291(Print) 2288-2790(Online)한국자원공학회

The construction of an accurate velocity model is one of the most important tasks in seismic data processing for hydrocarbon exploration. Because deep neural networks have garnered significant attention in the field of geophysics recently, studies have been performed to predict velocity models using regular convolutional neural networks. Herein, we propose a neural network with depthwise separable convolutional layers and an encoder–decoder structure to construct a velocity model. This network is trained using a supervised learning approach, and we predict P-wave velocity models from time-domain wavefields. In this network structure, depthwise separable convolutions perform spatial-oriented convolutions independently for each input channel. These depthwise separable convolutions can improve network performance while significantly reducing the number of model parameters as compared with regular convolutions. Synthetic velocity models generated for training contain various geological features, including folds, faults, and salt-dome structures. We compare a network with depthwise separable convolutions and a network with regular convolutions based on the same training conditions and hyperparameters. Experiments demonstrate that the network with depthwise separable convolutions is more efficient than the network with regular convolutions for constructing a seismic velocity model.