Adaptive learning 3D gravity inversion for salt-body imaging

We have developed an iterative scheme for inverting gravity data produced by salt bodies with density contrasts relative to the sediments varying from positive to negative, crossing, in this way, the nil zone. Our inversion method estimates a 3D density-contrast distribution, through a piecewise constant function defined on a user-specified grid of cells. It consists of two nested iterative loops. The outer loop uses an adaptive learning strategy that starts with a coarse grid of cells, a set of first-guess geometric elements (axes and points) and the corresponding assigned density contrasts. From the second iteration on, this strategy refines the grid and automatically creates a new set of geometric elements (points only) and associated density contrasts. Each geometric element operates as the first-guess skeletal outline of a section of the salt body to be imaged. The inner loop estimates the 3D density-contrast distribution for the grid of cells and for the set of geometric elements defined in the outer loop. The outer loop allows for easy incorporation of prior geologic information about the lithologic units and automatic evolution of the prior information. The inner loop forces the estimated density contrast of each cell to be close either to a null or to a non-null prespecified value. The iteration stops when the geometries of the estimated salt bodies are invariant along successive iterations. We apply our method to synthetic gravity data produced by a homogeneous salt body embedded in heterogeneous sediments. We tested two geologic hypotheses about the real gravity data from Galveston Island salt dome, USA. In the first, the estimated salt body attains a maximum bottom depth of 5 km, whereas in the second hypothesis, it is shallower and discloses an overhang. Both solutions fit the data and are feasible geologically, so both hypotheses are acceptable.