3-D magnetotelluric, seismic and potential field joint inversion via joint sparsity regularization
Abstract
J oint inversion of heterogeneous datasets is gradually becoming a standard tool of applied geophysics. Linking geophysical methods through a consistent physical theory, while possible in some cases, is often prohibitive both theoretically and computationally, one common example being 3-D magnetotelluric (MT) and seismic inversion on crustal scale. A simplified approach that proved to be highly efficient is to promote structural similarity between the inverted models, which is essentially equivalent to assuming that the physical properties are not completely independent. We develop a framework for regional-scale 3-D joint inversion of MT, gravity, magnetic and seismic traveltime data. The structural similarity constraint is imposed via joint sparsity regularization. Namely, we consider joint total variation and joint minimum gradient support functionals that act as stabilizers while promoting joint sparsity of the individual models' gradients, meaning that the preferred individual models tend to have edges at the same positions. This approach has several advantages over the standard cross-gradient coupling: there are less nonlinearities in the objective function, a single term of the objective function provides both regularization and coupling for any number of methods, resulting in only one hyperparameter (analogous to the classical regularization parameter) per method. This makes the proposed inversion framework robust and easily expandable. Inversion is formulated on a rectangular mesh, forward and adjoint problems are solved by finite-difference methods and the objective function is minimized by limited-memory BFGS or nonlinear conjugate gradient method. Performance is demonstrated on synthetic datasets. Our results show that the joint sparsity regularization helps to suppress individual-inversion artifacts by combining complimentary sensitivities, which is essential for reconciling the models before subsequent geological / thermochemical interpretation.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.S53D0487M
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1942 Machine learning;
- INFORMATICS;
- 3260 Inverse theory;
- MATHEMATICAL GEOPHYSICS;
- 3275 Uncertainty quantification;
- MATHEMATICAL GEOPHYSICS