Abstract
We present GEMMIE, a new open-source magnetotelluric (MT) three-dimensional inversion solver designed to handle large datasets with respect to survey area, period range, and number of observations. The main methodological innovations introduced in GEMMIE are: 1) a novel model parametrization strategy – the so-called non-local parametrization – which helps suppress artefacts near observation sites and significantly accelerates convergence to the inverse solution (by up to several times); 2) utilization of the recently developed version of the quasi-Newton iterative optimization method that exploits the structure of the regularized inverse problem and effectively eliminates the need for the additional iterations during line search; 3) the modelled fields interpolation technique that enables proper inverting data across a large number of periods preserving the integrity of the observed responses and their associated error estimates. The forward problem engine is inherited from the authors’ prior work and is based on a modern implementation of the volume integral equation approach, demonstrating near-linear scalability across thousands of computational cores. The workability of the presented solver and the efficacy of the proposed techniques are confirmed by validation on a synthetic dataset and benchmarked against results from inverting real data, obtained using a fundamentally different inverse solver based on the finite element method.