Semi-supervised intelligent inversion from prestack seismic attributes guided by geophysical prior knowledge

Abstract

Supervised deep learning methods currently used for prestack parameter prediction are suffered from the problem of limited training samples. The lack of clear physical meanings for deep learning models also makes prediction results unreliable. To address these issues, we proposed a geophysical prior knowledge guided semi-supervised (GPKGS) deep learning framework for amplitude-versus-angle (AVA) inversion. Based on prior physical knowledge, the prestack seismic data are decoupled into prestack seismic attribute data of the elastic parameters. Meanwhile, according to the prestack seismic attribute data, constructing the new forward models corresponding to each elastic parameter. The intelligent inversion framework is built based on the constructed forward models. This reduces the dependence of the framework on training data. This GPKGS framework preserves the physical procedure of AVA inversion, making intelligent inversion results reliable. The framework contains three branch networks for each elastic parameter. Each branch network contains an inversion neural network (INN) and a forward neural network (FNN). The INN can invert the prestack seismic attribute data into elastic parameters, which corresponding to inversion process. The FNN convert the obtained elastic parameters into synthetic prestack seismic attribute data, which corresponding to forward process. To ensure a reliable training process, the difference between the prestack seismic attribute data and the synthetic data are used to train the framework supervised by well log data. In addition, to obtain more stable results, at prediction stage, the prior information is introduced to help the FNN update the elastic parameters output by INN. The Marmousi2 model and a deep carbonate data are used to test the proposed framework. We find that the intelligent inversion results of the proposed network perform well at the situation of few training data.