Non-linear Conjugate Gradient Inversion For2.5D Airborne Transient Electromagnetic Data

The method of airborne electromagnetic (AEM) surveys usually collect tens of thousands of in-line data, implying a need for much more efficient data interpretation software than grounded-based EM methods. Despite of some differences of configuration, the numerical techniques of forwarding and inverting of AEM data is actually the same as the grounded-based controlled-source EM methods (CSEM). In this thesis, I devoted myself to developing and optimizing inversion codes of AEM based on available forwarding codes, which serves as an effort to better the level of interpretation for geophysical EM data in China.After briefly reviewing some recent progress of AEM inversions, I discussed the accuracy and efficiency of forwarding-modelling quantitatively and responses from a few simple, local conductivity models qualitatively. Numerical testing results of transient EM (TEM) responses with different square-loop source sizes but same current intensity indicates that computational accuracy of vertical magnetic field Hz is apparently different among different time samplings, especially worse in the later time. It suggests that one cannot guarantee a good calculation result by improving the accuracy of only one numerical technique whereas there are in fact several to be considered. Illustration on efficiencies of the forwarding codes demonstrates a possibility to invert2D TEM data in some scales if not large-scale data.I presented formula used for computing first-order derivative of EM responses (the so-called sensitivity) through the concept of adjoint field and proven its feasibility to invert2D/3D EM data. Numerical results of sensitivity for simple conductivity models clearly shows that time-domain sensitivity distributions in space has obvious similarities from that of frequency-domain in terms of decreasing rapidly from source to distant position in space, although the rate of decreasing in space depends on time samplings. Comparison of1D and2D sensitivity distribiution shows that it is feasible to approximate2D sensitivity using1D sensitivity. The method of choosing objective function to be optimized is illustrated in details and three commonly used optimization methods, namely GN,GN-CG and NLCG, are reviewed. Here, NLCG is chosen to invert synthesized TEM data which were calculated from a conductor embedded in a less conductive half-space host. However, the iteration results show that all three tests of NLCG inversion failed to converge with implemented line-search algorithm and methods for determining regularization parameter λ. The reasons of failure can be summarized in two aspects, namely, insufficient accuracy of later time TEM responses and too small NLCG step length. Some problems not solved well or completely in this thesis are also addressed in the end, suggesting further related discussions.