Seismic inversion based on GAN joined with geophysical constraints for near-surface geology

Deep-learning algorithms have great potential in interpreting complex geo-datasets, e.g., unveiling near-surface geological structures and properties by seismic inversion with deep-learning methods. However, it is realized that the lack of geophysical constraints seriously affects the feasibility and accuracy of the inversion results. To solve and address the above problem, we designed an improved generative adversarial network (GAN) and added new geophysical constraints to form an optimistic seismic inversion method based on joint GAN and GC (Geophysical Constraints). The established optimistic seismic inversion includes using conventional seismic inversion to expand and optimize pseudo-label training data sets, adding inversion and forward loss terms to the deep learning algorithm, and reconstructing the loss function to form a new GC. We tested the proposed method on the Marmousi model the field data, and the results show that the subsurface structure and seismic impedance predicted by this method have higher reliability