Three-dimensional magnetotelluric inversion using an adaptive algebraic multi-resolution sampling approach
Jian Li, Jianxin Liu, Yasuo Ogawa, Rongwen Guo, Xulong Wang, Yongfei Wang, Keke Zhou, Jingdao Xu
Abstract
ABSTRACT Conventional 3D magnetotelluric (MT) inversion methods face substantial computational challenges when inverting large-scale data sets, demanding extensive processing time and computing resources. To address these limitations, we develop an advanced frequency-polarization joint parallel inversion framework for MT data that uses the nonlinear conjugate gradient (NLCG) method, integrating an adaptive algebraic multi-resolution sampling (AMRS) approach with a regularization technique. The AMRS approach adaptively discretizes modeling domains across different periods based on the normalized root-mean-square difference relative to the solution on the initial grid, which serves as a reference. The regularization technique ensures the conservation of current, which can significantly affect the efficiency of the forward-modeling problem and its adjoint when using iterative solvers, particularly at low frequencies. The reliability and efficiency of the regularization AMRS inversion are first examined by benchmarking it against the traditional NLCG inversion and the regularization inversion based on a two adjacent blocks model and a terrain model. Subsequently, the regularization AMRS inversion is applied to invert real MT data from the Cascadia region in the USA, demonstrating its capability to address real inverse problems.