An efficient decoupled 3-D axial anisotropic resistivity inversion for magnetotelluric data with OpenMP parallelization
Abstract We have developed two novel axial anisotropic inversion codes for magnetotelluric (MT) data: a full axial inversion and a decoupled axial inversion. Both codes are based on the data space Gauss–Newton inversion method used in the WSINV3DMT code, enhanced with OpenMP parallelization and Inte...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
SpringerOpen
2025-01-01
|
| Series: | Earth, Planets and Space |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s40623-024-02113-5 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract We have developed two novel axial anisotropic inversion codes for magnetotelluric (MT) data: a full axial inversion and a decoupled axial inversion. Both codes are based on the data space Gauss–Newton inversion method used in the WSINV3DMT code, enhanced with OpenMP parallelization and Intel MKL PARDISO direct solver for improved computational efficiency. The full axial inversion searches for three axial anisotropic resistivity elements $${\rho }_{xx}$$ ρ xx , $${\rho }_{yy}$$ ρ yy , and $${\rho }_{zz}$$ ρ zz . Based on our prior influence studies, we identified contrasting influence patterns: $${\rho }_{xx}$$ ρ xx has a strong influence on Z xy and Z yy , whereas $${\rho }_{yy}$$ ρ yy has a strong influence on Z yx and Z xx , while $${\rho }_{zz}$$ ρ zz has negligible impact on all responses. We, therefore, proposed a decoupled axial inversion. This novel technique splits the full axial inversion into two independent modes: $${\rho }_{xx}$$ ρ xx - and $${\rho }_{yy}$$ ρ yy -modes. The $${\rho }_{xx}$$ ρ xx -mode requires just the Z xy and Z yy responses to recover $${\rho }_{xx}$$ ρ xx , while the $${\rho }_{yy}$$ ρ yy -mode requires the other two responses. By splitting the inversion, each mode requires half the data parameters and reduces model parameters to one-third compared to the full axial inversion. The substantial parameter reduction is the key factor leading to significantly faster processing times and lower memory requirements. Both codes were validated with the synthetic model. Utilizing OpenMP and direct solvers, the full axial inversion processed 256 MT stations in 65 min per iteration or a total of 195 min for three iterations to converge. The decoupled inversion achieved significantly faster processing, requiring just 32 min per iteration or a total of 96 min to finish the inversion. In addition, the decoupled inversion requires just a fraction of the memory used in the full axial inversion. This makes it practical to operate even on the standard personal computers of current technology. We further applied the decoupled inversion to real MT data acquired in the Northwestern Pacific Ocean. The resulting 3-D inverted $${\rho }_{xx}$$ ρ xx and $${\rho }_{yy}$$ ρ yy models support the existence of the anisotropy occurring at 60–120 km depth beneath the ocean floor, in agreement with previous studies. Graphical abstract |
|---|---|
| ISSN: | 1880-5981 |