An efficient data space conjugate gradient Occam's method for three-dimensional magnetotelluric inversion

Abstract

In this paper, we start with the implementation of a data space conjugate gradient (DCG) method for 3-D magnetotelluric (MT) data. This code will be referred to as WSDCG3DMT. It is an extension of the 2-D method previously developed. Several experiments on both synthetic and real data sets show that WSDCG3DMT usually needs more computational time than the data space Occam's inversion (OCCAM) for which the corresponding code is referred to as WSINV3DMT. However, the memory requirement of WSDCG3DMT is only a fraction of that of WSINV3DMT. Based on the knowledge gained from several studies of both codes, we have created a new hybrid scheme called the DCG Occam's inversion (DCGOCC) and the corresponding code, WSDCGOCC3DMT, from combining aspects of the OCCAM and DCG methods. As with OCCAM, the DCGOCC method divides the inversion into two phases. In Phase I the misfit is brought down to a desired level. In Phase II unnecessary structures are smoothed out. Because its mathematical basis is of a similar form to that of DCG, its memory requirement is similarly low but more stable. However, DCGOCC is significantly faster than both methods. We demonstrate the computational performances with comparisons of all three methods with both synthetic and EXTECH field data sets.