Abstract
The Spatio-temporal sequence residual modeling neural network (STSRM-net) is built to address the geophysical problem of obtaining P-impedance of the subsurface from the zero-offset seismic data and initial model of P-impedance. The initial model is used as the initial value of STSRM-net, and the modification value of the initial model and the residual between the modified initial model and true P-impedance are learned during the training. The Spatio-temporal characteristics of the data could be fully dug by STSRM-net. Limited by the inadequacy of labeled data, the semi-supervised learning inversion framework is constructed to train the STSRM-net. The performance of the STSRM-net and some deep learning inversion methods are compared on the synthetic dataset. The test results indicate that STSRM-net has higher inversion accuracy, stronger continuity, and better anti-noise performance. In addition, the robustness experiments indicate that the STSRM-net has excellent fault tolerance to the initial model. Finally, the STSRM-net is applied to the field data and compared with the sparse pulse inversion method, which also proves high inversion resolution, inversion accuracy and robustness to the choice of training wells of the proposed method.
NSTL