Advancements in an efficient, massive, and reality-oriented three-dimensional inversion framework for frequency-domain controlled-source electromagnetic data

This study introduces a novel and scalable 3-D anisotropic inversion framework for efficiently imaging controlled-source electromagnetic (CSEM) data in complex and realistic geological environments. The inversion framework first integrates a forward modeling engine using a tetrahedron-based vector finite-element approach to represent arbitrarily complex geometries accurately. Then, to ensure time and memory efficiency, a robust and well-convergent FGMRES iterative solver preconditioned by the auxiliary-space Maxwell solver (AMS) is employed to solve the large-scale forward and adjoint problems. The L-BFGS algorithm is chosen for minimizing the inverse objective function due to its advantage of requiring only one adjoint problem per iteration. In addition, we developed a double-level parallel scheme to perform the inversion procedure in parallel across sources, frequencies, and meshes for further efficiency gain. To avoid overparametrization for inversion, we adopt the tetrahedral longest-edge bisection technique to construct decoupled forward and inverse meshes, facilitating fast and accurate parameter mapping between different meshes. Finally, two synthetic models with undulating topography are tested to validate our inversion scheme, which shows high efficiency and well-scalability for massive inversion problems with millions of unknowns.