A framework for coupled physics-deep learning inversion and multiparameter joint inversion

A new approach to the inversion and joint inversion of geophysical data is described. We take advantage of the domains of local optimization and of the machine learning (ML) or deep learning (DL) technique to generate efficient optimization schemes to reduce uncertainties in the model parameter estimations, exploit the image segmentation capability of DL techniques, and guarantee compliance with the requirement of physics for the wave propagation. The domains of physics driven (Phy) optimization, based on data misfit functionals, and of DL optimization, based on model misfit (loss), are coupled by multiple penalty functions imposed on the common model term of the physical domain such as performed in a joint inversion approach. The procedure is complemented by network retraining with partial inversion results to augment the network knowledge base and enable more physics-oriented DL predictions. After several iterations, the procedure tends to converge to models satisfying both physics and DL optimization schemes by providing at the same time better resolution and accuracy in parameter estimation. The developed method is demonstrated on synthetic and field transient EM data.