Crosswell electromagnetic inversion using integral and differential equations

Abstract

The crosswell electromagnetic (EM) inverse problem is solved with an integral‐equation (IE) formulation using successive Born approximations in the frequency domain. Because the inverse problem is nonlinear, the predicted fields and Green’s functions are continually updated. Updating the fields and Green’s functions relates small changes in the predicted data to small changes in the model parameters through Fréchet kernels. These fields and Green functions are calculated with an efficient 3-D finite‐difference solver. Since the resistivity is invariant along strike, the 3-D fields are integrated along strike so the 2-D kernels can be assembled. At the early stages of the inversion, smoothing of the electrical conductivity stabilizes the inverse solution when it is far from convergence. As the solution converges, this smoothing is relaxed and more effort is made to reduce the data misfit. Bounds on the conductivity are included in the solution to eliminate unrealistic estimates. The robustness of the inversion scheme has been demonstrated with synthetic and field data that are underdetermined from the standpoint of the smooth models being sought. Two synthetic examples with added Gaussian noise were considered, including data arising from an IE solver. This IE solver is different from the one embedded in the inversion algorithm and has provided a stronger check on the scheme. The synthetic examples show it is more difficult to reconstruct a target’s conductivity than its geometry at a single frequency. The inversion scheme has been successfully tested using data collected at the Richmond‐field site near Berkeley, California, where it has imaged a salt water plume injected into the interwell region. The data in this experiment consisted of two sets of measurements, taken before and after the injection of 50 000 gallons of 1 Ωm salt water. Findings show that underdetermined inversion using small amounts of field data can be sufficient to produce useful, but smoothed, maps of the conductivity. The data in this instance need be only single frequency and single component.