Transformer and Convolutional Hybrid Neural Network for Seismic Impedance Inversion

The inversion of elastic parameters especially P-wave impedance is an essential task in seismic exploration. Over the years, deep learning methods have made significant achievements in seismic impedance inversion, and convolutional neural networks (CNNs) become the dominating framework relying on extracting local features effectively. In fact, the elastic parameters temporal correlation consists of local and global characteristics, with the latter as a general trend in vertical direction due to gravity and diagenesis (vertical mechanical compression). Therefore, considering the excellent performance in capturing global dependencies of Transformer, we design an improved transformer encoder, a transformer and convolutional hybrid neural network (trans-CNN), for seismic impedance inversion. The designed network not only has the ability of transformer capturing global features with the facilitation of parallel computing but also the advantage of extracting local features of CNNs. With sparse well log data as labels, it can infer the absolute impedance from seismic data without an initial model. We also devise a relative time interval prediction self-supervised task to assist the network in better extracting seismic data features without adding any labels. Therefore, a multitask framework composed of self-supervised and supervised learning is used to train the network. We first conduct experiments on the Marmousi2 and overthrust model. The prediction profiles show that the proposed trans-CNN has better inversion and transfer learning ability than several comparable networks. We then test the proposed network on a field data, the experiments further suggest that trans-CNN can obtain stable inversion results with better horizontal continuity and high vertical resolution.