A physics-augmented deep learning method for seismic data deblending Available to Purchase

In order to deliver affordable high-density seismic imaging, in recent years, simultaneous-source acquisition has gained momentum in the design of seismic reflection surveys. This high-productivity strategy relies on accurate deblending algorithms to eliminate the interference caused by simultaneous sources during the survey. Deep learning has shown its power in recognizing structural features in images and becomes a powerful tool for solving various seismic imaging problems. However, the training process of a deep learning model is not trivial. Particularly, it is difficult to obtain an adaptive model when there is less resemblance between the training and testing dataset. In this work, we proposed a physics augmented deep learning method, which is adaptive to various geological environments, to solve realistic deblending problems. A deep convolutional neural network is constructed to separate the signals from different sources and extract the coherent signals from the targeted source in a transform domain. Our method employs an iterative workflow to build representative training sets and adaptively update the model without collecting massive amount of labeled data. Testing results demonstrate the efficiency of the proposed approach and highlight the adaptiveness of the physics augmented deep learning workflow.