 |
| Titel |
Magnetotelluric Inversion in a 2D Anisotropic Environment |
| VerfasserIn |
E. Mandolesi, A. G. Jones |
| Konferenz |
EGU General Assembly 2012
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 14 (2012) |
| Datensatznummer |
250070975
|
|
|
|
|
|
| Zusammenfassung |
| Abstract
In recent years several authors have proposed algorithms to perform
magnetotelluric (MT) inversion in a 3D environment. The development of high
performance computer (hpc) machines allows the solution of these inverse
problems in a reasonable time, nevertheless the solution of a 3D problem remains
at the present extremely challenging. Moreover it is proofed that any magnitude
of anisotropy possibly present in the subsurface conductivity can be modeled by a
sufficient dense discretization of a 3D isotropic domain, keeping the recognition of
intrinsically anisotropic bulks virtually impossible for a 3D code. These arguments
convinced us to develop a 2D inverse code able in assessing anisotropy and running
in an affordable time, testing several scenario for the same dataset in the same
time in which a 3D inversion code produces its first model. In this work we report
results from synthetic tests we performed.
MT inverse problem is challenging because of several reasons. It is both
highly non-linear, ill-conditioned and suffers of a severe non-uniqueness of the
solution, therefore we developed an inversion algorithm based on the classic
Levenberg-Marquardt (LM) strategy, minimizing the objective function
(--G-(m)–d)2
Ï(m )= Ïă +λaLa +λsLs
in which m is the model, G the forward operator, d the data vector, L* the
regularization matrix and λ* the trade-off parameter for respectively anisotropy
and structure.
Usually LM method is used for medium size problems, mainly because it
requires the explicit computation and storage of the Jacobian matrix J and the
explicit knowledge of the product JTJ. To compute the Jacobian it has proofed
that the electrical reciprocity theorem is a valuable tool, allowing to compute the
full Jacobian with the evaluation of one forward problem per station in spite of one
forward problem per parameter as usually done with the finite-difference method.
Moreover the computation of the forward response can be easily performed in
parallel, due the mutual independency of the different spectral components, storing
the Jacobian in a distribute machine memory and solving at the same time the
problem of the huge memory requirements used to store the product JTJ and
speeding up the whole process.
We performed tests on the simple synthetic model released with the code from
Pek and Santos [2004]: an 84Ã100 cells grid, grouped in 3 up to 20 blocks sharing
the same conductivities. Results proof the capacity of the algorithm in recovering
the subsurface structure with good precision, reaching an RMS of the magnitude
10-5 for the 20 block case without the use of regularization. More tests will be
presented and results highlighted. |
|
|
|
|
|
|