arxiv: 2602.13690 · v2 · submitted 2026-02-14 · 💻 cs.LG

Recognition: no theorem link

Physics Aware Neural Networks: Denoising for Magnetic Navigation

Aritra Das (1) , Yashas Shende (1) , Muskaan Chugh (1) , Reva Laxmi Chauhan (1) , Arghya Pathak (1) , Debayan Gupta (1) ((1) Ashoka University)

Authors on Pith no claims yet

Pith reviewed 2026-05-15 22:10 UTC · model grok-4.3

classification 💻 cs.LG

keywords magnetic navigationphysics-informed neural networksdenoisingE(3) equivariancedivergence-free fieldsvector potentialtime series

0 comments

The pith

Neural networks enforcing divergence-free magnetic fields and E(3) equivariance denoise aircraft noise more accurately for navigation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper aims to show that incorporating two physical constraints into neural networks substantially improves their performance in denoising magnetic sensor data corrupted by aircraft-induced noise. A sympathetic reader would care because reliable magnetic navigation offers a backup when GPS signals are unavailable or jammed. The first constraint forces the output magnetic field to be divergence-free by having the network predict a vector potential and taking its curl. The second enforces correct transformation under rotations and translations using spherical harmonic tensor products. Experiments on synthetic data demonstrate that these constraints boost accuracy and physical consistency across several architectures, with the continuous-time Contiformer performing best.

Core claim

The central claim is that physics-aware neural networks, which output a vector potential A such that the magnetic field B equals the curl of A to satisfy div B = 0, and employ E(3)-equivariant layers based on spherical harmonics tensor products, learn to remove stochastic aircraft magnetic noise more effectively than unconstrained networks or classical methods. This is demonstrated through ablation studies on synthetic magnetic time series generated from the World Magnetic Model and conditional GANs, where the constrained models show superior predictive accuracy and adherence to physical laws.

What carries the argument

The combination of a neural network outputting a vector potential whose curl gives the divergence-free magnetic field, paired with spherical harmonic representations for E(3)-equivariant tensor products that ensure correct behavior under position and orientation changes.

Load-bearing premise

Enforcing the divergence-free condition and E(3) equivariance through the specified network outputs and layers will reduce the impact of stochastic aircraft noise in real magnetic measurements, beyond just the synthetic cases tested.

What would settle it

A direct comparison of denoising error and navigation accuracy between the physics-constrained models and baseline models on real airborne magnetic survey data collected under varying flight conditions; lack of improvement would falsify the practical benefit.

read the original abstract

Magnetic-anomaly navigation, leveraging small-scale variations in the Earth's magnetic field, is a promising alternative when GPS is unavailable or compromised. Airborne systems face a key challenge in extracting geomagnetic field data: the aircraft itself induces magnetic noise. Although the classical Tolles-Lawson model addresses this, it inadequately handles stochastically corrupted magnetic data required for navigation. To handle stochastic noise, we propose using two physics-based constraints: divergence-free vector fields and E(3)-equivariance. These ensure the learned magnetic field obeys Maxwell's equation and that outputs transform correctly with sensor position and orientation. The divergence-free constraint is implemented by training a neural network to output a vector potential A, with the magnetic field defined as its curl. For E(3)-equivariance, we use tensor products of geometric tensors represented via spherical harmonics with known rotational transformations. Enforcing physical consistency and restricting the admissible function space acts as an implicit regularizer that improves spatiotemporal performance. We present ablation studies evaluating each constraint alone and jointly across CNNs, MLPs, LTCs, and Contiformers. Continuous-time dynamics and long-term memory are critical for modelling magnetic time series; the Contiformer, which provides both, outperforms existing methods. To mitigate data scarcity, we generate synthetic datasets using the World Magnetic Model (WMM) and time-series conditional GANs, producing realistic, temporally consistent magnetic sequences across varied trajectories and environments. Experiments show that embedding these constraints significantly improves predictive accuracy and physical plausibility, outperforming classical and unconstrained deep learning approaches. Acknowledgement: This work was done in collaboration with Dirac Labs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows gains from physics constraints in neural magnetic denoising on synthetic data, but real-world aircraft noise handling is untested.

read the letter

The paper's central idea is to use a neural network that outputs a vector potential whose curl gives a divergence-free magnetic field, combined with spherical-harmonic tensor products to enforce E(3) equivariance. They integrate this into a Contiformer for continuous-time modeling of magnetic time series and compare it to other architectures. This approach is new in its specific application to aircraft magnetic noise removal. The ablations demonstrate that the constraints improve performance and plausibility over unconstrained versions across CNNs, MLPs, LTCs, and the Contiformer. Using the World Magnetic Model and cGANs to create synthetic datasets is a practical step given the difficulty of obtaining real labeled data. The main concern is that all results are on synthetic data. The stress-test note is right that the generators may miss some stochastic elements of real platform-induced noise, so the outperformance might not hold when applied to actual flight logs. The abstract does not include quantitative details like specific error metrics or variance, which makes it harder to judge the strength of the improvements. This work is for researchers in physics-informed machine learning, particularly those applying it to navigation or geophysical signals. It offers a clear example of embedding Maxwell's equations and symmetry constraints into a model. I would bring it to a reading group for discussion on the synthetic-to-real gap. It is worth peer review because the architectural contributions are concrete and the experiments are structured, even though real data tests would be needed to fully support the navigation claims.

Referee Report

3 major / 2 minor

Summary. The manuscript presents Physics Aware Neural Networks for denoising magnetic data in airborne navigation. It enforces two physics-based constraints: divergence-free magnetic fields by having the network output a vector potential A such that B = curl(A), and E(3)-equivariance using tensor products of spherical harmonics. These are applied to architectures including CNNs, MLPs, LTCs, and Contiformers, evaluated on synthetic magnetic time series from the World Magnetic Model and cGANs. The central claim is that these constraints improve predictive accuracy and physical plausibility over classical and unconstrained approaches, with Contiformers performing best due to continuous-time dynamics.

Significance. The results, if they generalize, would be significant for the field of magnetic navigation as they demonstrate how embedding Maxwell's equations and symmetry constraints can serve as effective regularizers for noisy sensor data. The use of synthetic data generation to address scarcity is practical, and the architectural (rather than loss-based) enforcement of constraints is a clean approach that does not introduce circularity. The emphasis on long-term memory models for time series is well-motivated.

major comments (3)

[Experiments] Experiments section: The ablation studies across CNNs, MLPs, LTCs, and Contiformers lack quantitative error bars, exact data exclusion criteria, and verification on held-out real (non-synthetic) flights, undermining the claim that improvements hold for stochastic aircraft-induced noise.
[Method] Method section: While the divergence-free constraint is implemented via vector potential output, the manuscript does not specify how the curl operation is computed in the discrete time-series setting or its impact on gradient flow during training.
[Evaluation] Evaluation: All reported gains are on trajectories synthesized from WMM and cGANs; the central claim requires but does not provide evidence that the constraints mitigate real-world platform noise rather than only synthetic artifacts.

minor comments (2)

[Abstract] Abstract: The abstract mentions collaboration with Dirac Labs but this is not expanded in the main text; consider adding a brief statement on contributions.
[Notation] Notation: Clarify the distinction between the magnetic field B and the vector potential A in the equations to avoid confusion for readers unfamiliar with the Tolles-Lawson model.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive comments, which help clarify the scope and limitations of our work on physics-aware neural networks for magnetic denoising. We address each major comment below and indicate revisions where the manuscript will be updated.

read point-by-point responses

Referee: [Experiments] Experiments section: The ablation studies across CNNs, MLPs, LTCs, and Contiformers lack quantitative error bars, exact data exclusion criteria, and verification on held-out real (non-synthetic) flights, undermining the claim that improvements hold for stochastic aircraft-induced noise.

Authors: We agree that error bars and data criteria should be reported. In the revised manuscript we will add standard deviations over 5 independent runs for all ablation metrics and explicitly state the exclusion criteria (e.g., trajectories with >20% missing samples or unrealistic acceleration spikes). Regarding real flights, our study deliberately uses synthetic data generated from the World Magnetic Model and cGANs precisely because large-scale, labeled real-world airborne magnetic datasets with known platform noise are not publicly available; we therefore cannot provide held-out real-flight verification at this time. revision: partial
Referee: [Method] Method section: While the divergence-free constraint is implemented via vector potential output, the manuscript does not specify how the curl operation is computed in the discrete time-series setting or its impact on gradient flow during training.

Authors: We will expand the Method section with a precise description: the curl is obtained via automatic differentiation (torch.autograd.grad or equivalent) on the network-output vector potential A, using central finite differences on the discrete time grid for the spatial components. Because the operation is fully differentiable, gradients flow end-to-end without additional approximation; we will include pseudocode and a short gradient-flow diagram in the revision. revision: yes
Referee: [Evaluation] Evaluation: All reported gains are on trajectories synthesized from WMM and cGANs; the central claim requires but does not provide evidence that the constraints mitigate real-world platform noise rather than only synthetic artifacts.

Authors: The physics constraints (divergence-free via curl(A) and E(3)-equivariance) are architecture-level and independent of the data source; they restrict the hypothesis space to physically admissible fields, which we expect to generalize. The cGAN component was trained to reproduce statistical properties of real platform noise observed in limited flight logs. Nevertheless, we acknowledge that direct empirical evidence on real flights is absent and have added a dedicated limitations paragraph discussing this point and the reliance on synthetic data. revision: partial

standing simulated objections not resolved

Verification on held-out real (non-synthetic) flights, as sufficiently large public datasets with ground-truth platform noise are unavailable.

Circularity Check

0 steps flagged

No significant circularity; physics constraints imposed architecturally

full rationale

The paper's central mechanism outputs a vector potential A and defines B = curl(A) to enforce div B = 0 by the vector calculus identity, and implements E(3)-equivariance via spherical-harmonic tensor products using known rotation matrices. These are fixed architectural choices, not parameters fitted to the target magnetic time series or derived from the paper's own predictions. Ablation results and comparisons to CNN/MLP/LTC/Contiformer are empirical measurements on held-out synthetic trajectories; they do not reduce to a quantity defined by the authors' fitted values. No self-citations, uniqueness theorems, or renamings appear in the provided text that would make the claimed improvement tautological. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on two standard domain assumptions from electromagnetism and group theory; no new entities are postulated and no free parameters are explicitly fitted beyond ordinary network weights.

axioms (2)

domain assumption Magnetic fields satisfy ∇ · B = 0 (Maxwell's equations)
Invoked to justify outputting vector potential A and defining B = ∇ × A
domain assumption The magnetic field transforms correctly under E(3) group actions (rotations and translations)
Used to motivate spherical-harmonic tensor-product layers

pith-pipeline@v0.9.0 · 5626 in / 1399 out tokens · 34778 ms · 2026-05-15T22:10:29.914899+00:00 · methodology

Physics Aware Neural Networks: Denoising for Magnetic Navigation

Core claim

What carries the argument

Load-bearing premise

What would settle it

discussion (0)