A flexible regularization for magnetic vector inversion with petrophysical, structural, and sparsity constraints
Abstract
Magnetic vector inversion (MVI) is widely used because it can account for both induced and remanent magnetization by inverting for a magnetization vector model. This flexibility comes at the cost of nonuniqueness, since the number of unknowns is three times that of scalar inversion, and prior information can be required to make the result geologically useful. Recent work has applied sparse norms to the vector model, but many of the other tools for incorporating prior information within scalar inversions remain either unavailable for a vector model or unavailable without introducing artifacts or significant nonlinearity. We introduce a vector regularization, formulated in Cartesian coordinates, that generalizes existing sparse Cartesian formulations. We formulate the smallness term to admit a Cartesian reference model, from susceptibility and oriented remanence, and a magnitude reference model, from unoriented remanence or total magnetization. Because a nonzero magnitude reference model makes the Hessian of the smallness term indefinite, we derive its eigendecomposition analytically and modify the eigenvalues so that standard conjugate-gradient solvers can be applied. We formulate the smoothness term to penalize changes in both the direction and magnitude of magnetization. It is independent of the choice of Cartesian basis, so sparse norms can be applied without coordinate-aligned artifacts, and it admits arbitrary preferred smoothing directions through a rotated gradient operator. We demonstrate the method on synthetic data from a folded body with spatially variable magnetization direction, where each form of prior information improves the recovered model. We then apply it to airborne magnetic data over the DO-27 kimberlite, where incorporating petrophysical measurements yields a model more consistent with the measured properties and the geologic model inferred from drilling.
Related
- Bayesian Inversion Beyond the Generator's Range: Reducing Non-physical Constraints from Generative Neural Networks
- Conditional Climate–Cost Trade-Offs and Transport Break-Even Conditions for Waste Flexible-Film Recycling Under Feedstock Quality Constraints
- Binding mechanisms dictate distinct interactions of iron oxides with nanoplastics under structural feature constraints
- Relationship between soil electrical conductivity and water content: A multiscale petrophysical calibration
- A data-driven method coupling multiple physical constraints for correcting structural errors in groundwater contaminant transport models
BibTeX
@misc{weis2026flexible,
title = {{A flexible regularization for magnetic vector inversion with petrophysical, structural, and sparsity constraints}},
author = {John M. Weis and Lindsey J. Heagy},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.36368},
eprint = {2610.36368},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.36368},
}
RIS
TY - UNPB
TI - A flexible regularization for magnetic vector inversion with petrophysical, structural, and sparsity constraints
AU - John M. Weis
AU - Lindsey J. Heagy
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.36368
UR - https://arxiv.org/abs/2610.36368
AB - Magnetic vector inversion (MVI) is widely used because it can account for both induced and remanent magnetization by inverting for a magnetization vector model. This flexibility comes at the cost of nonuniqueness, since the number of unknowns is three times that of scalar inversion, and prior information can be required to make the result geologically useful. Recent work has applied sparse norms to the vector model, but many of the other tools for incorporating prior information within scalar inversions remain either unavailable for a vector model or unavailable without introducing artifacts or significant nonlinearity. We introduce a vector regularization, formulated in Cartesian coordinates, that generalizes existing sparse Cartesian formulations. We formulate the smallness term to admit a Cartesian reference model, from susceptibility and oriented remanence, and a magnitude reference model, from unoriented remanence or total magnetization. Because a nonzero magnitude reference model makes the Hessian of the smallness term indefinite, we derive its eigendecomposition analytically and modify the eigenvalues so that standard conjugate-gradient solvers can be applied. We formulate the smoothness term to penalize changes in both the direction and magnitude of magnetization. It is independent of the choice of Cartesian basis, so sparse norms can be applied without coordinate-aligned artifacts, and it admits arbitrary preferred smoothing directions through a rotated gradient operator. We demonstrate the method on synthetic data from a folded body with spatially variable magnetization direction, where each form of prior information improves the recovered model. We then apply it to airborne magnetic data over the DO-27 kimberlite, where incorporating petrophysical measurements yields a model more consistent with the measured properties and the geologic model inferred from drilling.
ER -