HR: 1340h
AN: GP33B-1254 [Abstracts]
TI: A Modular System for EM Inversion: Implementation of NLCG for 3-D Magnetotelluric Data
AU: * Kelbert, A
EM: anya@coas.oregonstate.edu
AF: COAS/OSU, 104 COAS Admin. Bldg., Corvallis, OR 97331-5503, United States
AU: Egbert, G D
EM: egbert@coas.oregonstate.edu
AF: COAS/OSU, 104 COAS Admin. Bldg., Corvallis, OR 97331-5503, United States
AB:
We are developing a general modular system for gradient based inversion of electromagnetic (EM) data. The
inversion code
has been designed using an object oriented approach,
with the highest levels of functionality independent of problem specifics. In particular, top level modules
manipulate abstract data objects
(such as data vectors, model parameters, EM field solutions) and methods
(e.g., data functionals, solvers), to implement a range of model and data space
gradient based search algorithms.
Previously, the system has been used as a test-bed for experimentation with new inversion algorithms, using the
2D magnetotelluric (MT) problem as a test case. Here we describe progress on using this general modular
system to develop a non-linear conjugate gradients inversion for 3D MT data. The kernel of this new inversion is a
3D staggered-grid finite difference forward solver. The very specific interfaces of the forward code, both with
respect to model parameters and source/boundary conditions,
facilitate incorporation into the general modular system.
The full adjoint (including sensitivity to boundary conditions) for the solver was developed as an intrinsic
part of the forward code, making application in a gradient based inversion scheme such as non-linear conjugate
gradients (NLCG) particularly straightforward.
Development of such an efficient, ultimately parallel, gradient-based inversion has been motivated by the newly
available, intrinsically 3D MT data sets such as those collected in the framework of EarthScope. We expect that
the modular approach will simplify development and testing of alternative parameterization and regularization
schemes, more readily allow inversion of novel data types (e.g., inter-station or magnetic gradient transfer
functions), and accelerate implementation of new more efficient search algorithms.
DE: 0639 Nonlinear electromagnetics
DE: 0644 Numerical methods
DE: 1515 Geomagnetic induction
DE: 3225 Numerical approximations and analysis (4260)
DE: 3260 Inverse theory
SC: Geomagnetism and Paleomagnetism [GP]
MN: 2007 Fall Meeting