Physics-Constrained Deep Ensemble for 3-D Magnetic Gradient Tensor Inversion With Uncertainty Quantification
IEEE Access · 2026 · 14 · 98861-98875
DOI: 10.1109/access.2026.3705583 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shengtao Xue und 6 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
To address the nonlinearity, ill-posedness, and non-uniqueness of conventional magnetic anomaly inversion, as well as the limited reliability characterization of existing deep-learning-based methods, this study proposes a three-dimensional probabilistic inversion method for magnetic gradient tensor data based on physical constraints and Deep Ensemble. Using U-Net++ as the backbone network, the proposed method maps six-component magnetic gradient tensor observations to three-dimensional voxel occupancy probabilities. During training, an observation-domain gradient-consistency constraint is introduced to enhance the physical plausibility and structural consistency of the inversion results. During inference, multiple independently trained sub-models are statistically aggregated to obtain voxel-wise predictive means and variances, thereby providing probabilistic inversion results and model-disagreement-based reliability information. Synthetic experiments show that the physical constraint improves the F1 score from 0.81 to 0.87 and the IoU from 0.76 to 0.80. Uncertainty-validation experiments show that the predictive variance is related to voxel-wise prediction error, with an AURC of 0.0044 under the 5% noise setting, and increases as the observation noise level rises. Field results from a controlled experiment demonstrate that the proposed method can recover the main spatial locations and overall geometric outlines of two coexisting ferromagnetic targets, while the predictive variance highlights target edges and locally disturbed regions. The proposed method provides a feasible route for physics-constrained probabilistic inversion of three-dimensional magnetic gradient tensor data.
Zitieren
APA
Xue, S., Li, Z., Fan, H., Wang, B., Li, J., Lan, J., & Xie, D. (2026). Physics-Constrained Deep Ensemble for 3-D Magnetic Gradient Tensor Inversion With Uncertainty Quantification. IEEE Access, 14, 98861-98875. https://doi.org/10.1109/access.2026.3705583
Vancouver
Xue S, Li Z, Fan H, Wang B, Li J, Lan J, et al. Physics-Constrained Deep Ensemble for 3-D Magnetic Gradient Tensor Inversion With Uncertainty Quantification. IEEE Access. 2026;14:98861-98875. doi:10.1109/access.2026.3705583
BibTeX
@article{Xue2026,
author = {Xue, Shengtao and Li, Zhining and Fan, Hongbo and Wang, Bo and Li, Jun and Lan, Jie and Xie, Deqian},
title = {Physics-Constrained Deep Ensemble for 3-D Magnetic Gradient Tensor Inversion With Uncertainty Quantification},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {98861--98875},
doi = {10.1109/access.2026.3705583}
}
EUVIMED