A stable and efficient attenuation compensation method based on physics-constrained deep learning

In this paper, a physics-constrained deep-learning-based attenuation compensation method is proposed. We propose a time-domain attenuation model and use it to constrain deep learning to realize unsupervised or semi-supervised attenuation compensation. Benefiting from the insensitivity of deep neural networks to noise and the laws of physics, the proposed method can achieve stable and accurate compensation even strong noise present in seismic data. Moreover, thanks to the efficiency of deep learning, this method significantly improves the efficiency of attenuation compensation. Both synthetic and field data experiments verified the effectiveness of the proposed method and demonstrated its advantages over common methods.