Abstract:
Inverse modelling is an ill-posed problem, and the regularisation technique is used to solve the inverse problem by biasing the solution with prior information. In the absence of prior information about the model, a constraint that seeks the smoothest model is enforced. However, a regularisation scheme that imposes information obtained from other geophysical methods has the potential to produce a more realisticmodel with sharp boundaries of anomalies. This thesis presents a study on the development of a structurally constrained inversion algorithmof controlled-source electromagnetic (CSEM) data. We have devised an efficient space-domain forward and gradient computation algorithm for CSEMdata. The space-domain simulation is achieved by imposing novel boundary conditions on the plane perpendicular to the strike direction that passes through the transmitter position. In this study, the boundary conditions for various transmitter types are derived using the symmetric/antisymmetric character of the electric and magnetic fields. For all the other boundaries, a homogenousDirichlet boundary condition is applied. The devised strategy facilitates efficient computation as one needs to discretise the space only on the side of the source position along the strike direction. Furthermore, the benchmarking experiment reveals that only six to eight grids are sufficient for discretisation in the strike direction for the accuracy required in geophysical data analysis. A Gauss-Newton optimisation-based inverse modelling algorithm is developed for two-dimensional (2D) CSEMdata inversion by employing the proposed forward modelling algorithm. The algorithmcan performinversion for the vertical transverse isotropy (VTI) subsurface model. The adjoint approach is used for the computation of the Jacobian matrix. A Gauss-Newton method is employed to calculate updated model parameters where the Hessian matrix is solved using a conjugate gradient solver. The developed
algorithm is tested for synthetic and real-field CSEMdata, and the inversion experiment agrees with the benchmarking of forward modelling, indicating that around eight are sufficient for discretisation in the strike direction. The comparison of the proposed algorithm with a published algorithm shows that our algorithmis at least one order faster in terms of computation time and requires lessmemory.
