Parallelizing the 2.5D airborne electromagnetic inversion program ArgunAir

Abstract

This thesis describes the development of a parallel version of the 2.5D airborne electromagnetic modelling and inversion program ArjunAir. The program uses a finite-element scheme to model the response of an earth with a 2D conductivity structure to a 3D electromagnetic source. The program uses a Gauss-Newton like iterative inversion algorithm, stabilized by singular value damping, to estimate the conductivity of a 2D depth section of the earth beneath an airborne electromagnetic survey line. The forward modelling code was parallelized and whenever possible, bottleneck routines were replaced by more efficient versions. Shared and distributed memory parallel versions of the ArjunAir forward solver were developed, with the shared memory version being incorporated into a modified ArjunAir inversion program based on the Levenberg-Marquardt algorithm. The new shared memory parallel ArjunAir inversion algorithm ran up to 8 times faster than the original algorithm when running with 8 threads, with speedup due both to parallelization and the use of more efficient sequential routines.