Deep Learning Decouples Magnetic Gradient Inversion and Quantifies Three-Dimensional Subsurface Uncertainty

A computational framework combining physics-constrained deep learning, analytical magnetization estimation, and Bayesian uncertainty modeling can accurately recover three-dimensional target geometry from two-dimensional magnetic gradient tensor measurements while characterizing spatial ambiguity. Developed by researchers Shengtao Xue, Zhining Li, Hongbo Fan, Bo Wang, Jun Li, Jie Lan, and Deqian Xie, the methodology addresses a persistent challenge in geophysical exploration: the mathematical coupling between subsurface geometry and target magnetization direction. When remanent magnetization causes total magnetization to deviate from the ambient geomagnetic field, traditional scalar susceptibility inversion creates distorted boundaries, shifted anomaly centers, and erroneous interpretations. Rather than requiring computationally intensive voxel-by-voxel vector inversions or relying on predefined magnetic directions, the new framework decouples geometric shape prediction from magnetization estimation, allowing rapid spatial reconstruction accompanied by localized reliability metrics.

Subsurface potential-field inversion from data collected across a single planar survey is an inherently ill-posed and under-determined mathematical problem. Measurements of the six-component magnetic gradient tensor provide richer spatial-gradient information than total field measurements alone, but they do not eliminate non-uniqueness. Standard magnetization vector inversion formulations solve this by assigning three independent Cartesian vector components to every discretized cell in a subsurface mesh, generating a high-dimensional optimization task that requires complex regularization and extensive iterative computation. Conversely, purely data-driven neural networks that attempt to jointly predict both high-dimensional target occupancy and directional magnetic properties often overfit to the specific parameter distributions present in their training datasets without verifying whether predicted shapes actually reproduce the measured field anomalies.

To overcome these limitations, the research team exploited the linear physical relationship governing magnetic forward responses when geometry is held fixed. For any defined three-dimensional target shape, the resulting magnetic gradient tensor fields scale linearly with the three orthogonal Cartesian components of a uniform equivalent magnetization vector. The researchers utilized this linearity to separate the non-linear task of geometric prediction from the linear task of magnetization parameter estimation. In the primary network pipeline, an encoder–decoder architecture based on U-Net++ receives normalized two-dimensional arrays representing the six independent tensor components—Bxx, Bxy, Bxz, Byy, Byz, and Bzz—and predicts a three-dimensional target occupancy probability grid. The equivalent uniform magnetization vector is then determined analytically through a stabilized linear least-squares solution, bypassing directional grid searches and reduction-to-the-pole preprocessing.

The structural backbone of the geometric predictor incorporates dense skip connections and nested sub-paths characteristic of U-Net++, which minimize the semantic discrepancy between multi-scale feature representations in the encoder and decoder stages. The encoder downsamples input tensor patterns to capture broad structural dimensions and global anomaly centers, while the decoder progressively restores spatial resolution to delineate sharp anomaly edges and local gradient variations. The resulting output layer transforms the decoded features into a multi-channel volume, which is passed through a sigmoid activation function to output voxel-wise occupancy probabilities bounded between zero and one. Training was conducted over 100 epochs with the Adam optimizer, utilizing a composite loss function that combines voxel-wise binary cross-entropy, a Dice loss to counteract severe foreground-to-background volumetric imbalance, and a physics-consistency loss evaluated in observation space.

Observation-space physical consistency is achieved by feeding the network’s predicted target geometry into a magnetic forward operator based on analytical rectangular prism formulations. The algorithm computes the forward response of the candidate geometry under three orthogonal unit magnetization vectors along the x, y, and z axes to construct a forward response matrix. Multiplying this response matrix by the analytically calculated least-squares magnetization vector generates a reconstructed six-component tensor dataset. The physics-consistency loss evaluates the relative root-mean-square discrepancy between the reconstructed observations and the raw input data. Because this physics loss dynamically adapts to the optimal uniform magnetization vector calculated for each predicted model, it enforces fidelity to physical potential-field equations without imposing rigid assumptions regarding true magnetization inclination or declination.

To quantify the structural ambiguity inherent in single-plane potential-field data, the researchers implemented a local, network-centered Bayesian posterior uncertainty formulation. Instead of attempting computationally prohibitive Markov chain Monte Carlo exploration across tens of thousands of voxel variables, the framework generates candidate geometric variations in logit space around the primary network prediction. Perturbations drawn from zero-mean Gaussian random fields are spatially smoothed using a three-dimensional Gaussian kernel to maintain spatial continuity before being mapped back to probability space. For each perturbed candidate geometry, the system analytically computes the optimal equivalent magnetization vector, evaluates the corresponding forward-modeling response against the measured tensor data, and assigns a likelihood-based posterior weight assuming Gaussian observational errors.

The weighted candidate models yield both a posterior mean occupancy probability distribution, which serves as the final three-dimensional inversion model, and a voxel-wise posterior variance distribution that highlights spatial uncertainty. Numerical experiments demonstrated that regions exhibiting elevated posterior variance coincide with target boundaries, transition zones, and areas affected by noise or weak potential-field responses. Stratified risk-coverage analyses on synthetic validation datasets confirmed a direct positive correlation between posterior variance and actual geometric prediction errors. Progressively removing voxels characterized by higher posterior variance consistently reduced the average error across the remaining model volume, establishing the posterior variance metric as a practical filter for isolating highly reliable structural interpretations.

Quantitative evaluations conducted on an extensive synthetic corpus comprising 81,000 forward-modeled samples—including block-, strip-, and plate-shaped anomalies—verified the operational stability of the decoupled architecture. Synthetic targets tested across vertical and oblique magnetization configurations demonstrated high geometric fidelity, with the analytical least-squares module recovering magnetization orientations with low angular errors, such as 1.29 degrees under vertical conditions and 3.24 degrees under oblique alignments. Robustness evaluations subjected the trained framework to target center depths ranging from 0.20 meters to 0.50 meters and observation noise levels up to 10 percent. Although geometric overlap metrics gradually decreased with deeper targets due to natural gradient attenuation, the network consistently identified core target positions without requiring structural retraining.

To evaluate the framework under practical operational conditions, the researchers conducted a controlled field experiment at a testing facility in Shijiazhuang, Hebei Province, China. The test utilized a cross-shaped sensor array containing five triaxial fluxgate magnetic sensors mounted on a non-magnetic frame with a 16-centimeter baseline, suspended 1.2 meters above ground level over a 1.2-meter by 1.2-meter survey plot. Two ferromagnetic targets were buried within the survey footprint: a disc-shaped target measuring 35 centimeters in diameter and 10 centimeters in thickness buried at a depth of 0.27 meters, and a block-shaped target measuring 25 centimeters by 35 centimeters by 10 centimeters buried at a depth of 0.30 meters. Unprocessed measurements were digitized via a 24-bit acquisition card, capturing complex superposition effects, instrument noise, and ambient background disturbances.

Inversion of the real-world survey measurements accurately delineated both targets despite substantial field noise and overlapping magnetic responses. The predicted center positions deviated from the known deployment coordinates by only 0.66 voxels for the disc target and 0.48 voxels for the block target, with both exhibiting depth recovery errors of 0.01 meters. Notably, the network had never encountered disc-shaped geometries during training, yet it successfully isolated the disc’s spatial boundaries, confirming strong generalization capacity. Forward modeling from the inverted posterior mean geometry reproduced the raw field observations with a mean root-mean-square error of 2469.56 nT/m, representing 8.59 percent of the maximum measured signal amplitude.

Computational efficiency benchmarks highlight the system’s operational viability for field data processing. On an NVIDIA GeForce RTX 4080 graphics processing unit, full supervised model training required approximately 2.7 hours, whereas deterministic single-sample inference executed in roughly 0.03 seconds. Evaluating fifty candidate samples during Bayesian uncertainty modeling completed in approximately 1.2 seconds per survey grid. Although multi-target tests revealed that severe magnetization contrast between adjacent bodies can degrade single-vector approximations, the integration of deep geometric extraction, analytical vector estimation, and localized uncertainty scoring establishes an effective computational pathway for processing ambiguous magnetic gradient tensor datasets.

Subject of Research: Three-dimensional magnetic gradient tensor inversion using deep learning, analytical magnetization estimation, and Bayesian uncertainty modeling

Article Title: 3D inversion of magnetic gradient tensor data integrating analytical magnetization-vector estimation and Bayesian uncertainty quantification

Article References: Xue, S., Li, Z., Fan, H., Wang, B., Li, J., Lan, J., & Xie, D. (2026). 3D inversion of magnetic gradient tensor data integrating analytical magnetization-vector estimation and Bayesian uncertainty quantification. Discover Computing, 29(1), Article 784. https://doi.org/10.1007/s10791-026-10729-2

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10729-2

Keywords: magnetic gradient tensor, 3D inversion, deep learning, U-Net++, magnetization vector inversion, Bayesian uncertainty quantification, geophysics, remanent magnetization, potential field, least-squares estimation, physics-constrained neural network, subsurface mapping, fluxgate magnetic sensors, Dice loss, synthetic modeling, forward modeling, ferromagnetic target detection, probabilistic inversion