Deep learning-enhanced multiphysics joint inversion
Yanyan Hu, Xiaolong Wei, Xuqing Wu, Jiajia Sun, Jiefu Chen, Jiuping Chen, Yueqing Huang
Abstract
In this paper, we design a framework to combine deep neural network (DNN) and the traditional separate inversion workflow together and improve the joint inversion result iteratively for multi-physics data. Different from conventional end-to-end networks, our proposed deep learning enhanced (DLE) joint inversion framework takes the separately inverted models, instead of data, as the inputs of the DNN, constructing structure similarity constraints during the learning process where no specific relationship between different property values is assumed in advance. Reliable recovered results are guaranteed by additional weighting and cooling strategy and well-defined iterations. The network can be easily extended to incorporate multi-physics without structural changes. Numerical examples demonstrate more accurate inverted property values and structural features of our DLE method compared to traditional separate inversions and cross gradient based joint inversion. Finally, excellent generalization abilities of the learning-based framework are validated by testing on datasets obtained by different sensing configurations or using divergent geological subsurface formations.