Physics-driven deep learning joint inversion
Daniele Colombo, Weichang Li, Diego Rovetta, Ernesto Sandoval‐Curiel, Erşan Türkoğlu
Abstract
Machine learning (ML) and specifically deep learning (DL) techniques applied to inversion problems are still a relatively new area of research which is appealing to geophysical applications. We developed a hybrid workflow combining the efficiency of physics-driven inversion with the power of data-driven DL based inversion. The two procedures are coupled by the model term. The method involves re-training of the network after each inversion iterations. The coupled inversion schemes are evolving and balancing each other to converge to a common model satisfying the data misfit criteria and the optimization of the DL network parameters at the same time. The benefits are related to the availability of a range of stochastically-sampled prior models for physics-driven inversion and of extended training sets for DL inversion. The coupled procedure is expected to reduce the non-uniqueness of the inversion results. The hybrid physics-driven DL inversion is formulated for general multiparameter joint inversions and it is demonstrated on a synthetic transient EM (TEM) dataset. Presentation Date: Tuesday, October 13, 2020 Session Start Time: 8:30 AM Presentation Time: 11:25 AM Location: 360C Presentation Type: Oral