🤖 machine learning

Physics Aware Neural Networks: Denoising for Magnetic Navigation

This paper proposes a physics-aware neural network framework for magnetic navigation that integrates divergence-free constraints and E(3)-equivariance to effectively denoise aircraft-induced magnetic interference, demonstrating that the Contiformer architecture combined with synthetic data generation significantly outperforms classical and unconstrained deep learning approaches.

Original authors: Aritra Das, Yashas Shende, Muskaan Chugh, Reva Laxmi Chauhan, Arghya Pathak, Debayan Gupta

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aritra Das, Yashas Shende, Muskaan Chugh, Reva Laxmi Chauhan, Arghya Pathak, Debayan Gupta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Navigating in a Magnetic Fog

Imagine you are flying a plane, but your GPS is broken or being jammed by an enemy. You can't see the ground, and you can't rely on satellites. How do you know where you are?

One solution is Magnetic Navigation. Think of the Earth's magnetic field like a giant, invisible fingerprint map. Every spot on Earth has a slightly unique magnetic "texture." If you have a sensitive compass, you can feel these tiny textures and figure out exactly where you are.

The Catch: The plane itself is made of metal, has engines, and carries electronics. These things create their own magnetic "static" or noise. It's like trying to listen to a quiet radio station while standing next to a loud, buzzing refrigerator. The compass on the plane hears the Earth's map plus the plane's own noise, making the signal useless for navigation.

The Old Way vs. The New Way

The Old Way (The "Math Filter"):
For decades, scientists used a classic math formula (the Tolles-Lawson model) to try to subtract the plane's noise. It's like using a basic noise-canceling headphone that works well for a steady hum but fails when the noise is chaotic and random. The paper says this old method isn't good enough for the messy, real-world noise planes actually face.

The New Way (The "Physics-Aware AI"):
The authors built a new type of Artificial Intelligence (a Neural Network) that doesn't just guess how to remove noise. Instead, they taught the AI the laws of physics that magnetic fields must follow. They forced the AI to act like a physicist, not just a statistician.

They added two "rules" to the AI's brain:

Rule 1: The "No Dead Ends" Rule (Divergence-Free)

The Physics: Magnetic field lines are like loops. They never start or stop at a single point; they always form a circle or go off to infinity. You can't have a magnetic "dead end" (a monopole).
The Analogy: Imagine a river. Water flows in a current; it doesn't just appear out of thin air in the middle of a field, nor does it vanish into a hole.
The Fix: The team programmed the AI so that it cannot output a magnetic field that has dead ends. If the AI tries to create a "dead end" (which is actually just noise), the math forces it to correct itself. It's like building a riverbed that physically prevents water from stopping in the middle of a field.

Rule 2: The "Shape-Shifting" Rule (E(3)-Equivariance)

The Physics: If you turn your compass or move it to a new spot, the magnetic field reading should change in a very specific, predictable way. If you rotate the sensor 90 degrees, the numbers should rotate 90 degrees, not scramble randomly.
The Analogy: Imagine holding a spinning top. If you turn your head, the top looks different, but it's still the same top. If you look at it from the left, it's a mirror image. If you look from the right, it's the other way. The AI needs to understand that the object (the magnetic field) stays the same, even if your view (the sensor) changes.
The Fix: The AI was built with a special "geometric brain" that knows how to rotate and move data correctly. It ensures that if the plane turns, the AI's prediction turns with it perfectly, rather than getting confused and producing garbage data.

The Secret Sauce: Time and Memory

The paper also noticed that magnetic noise isn't just a random spike; it's a slow, lingering mess.

Continuous Time: Real sensors don't take measurements at perfect, equal intervals (like a metronome). They jitter. The AI uses a "Neural ODE" (a type of math that handles time as a smooth flow, not a series of steps) to handle this jittery timing.
Long-Term Memory: Some noise (like heat buildup in the engine or electrical currents in the metal) lasts for a long time. The AI uses a special architecture (called a Contiformer) that has a "long-term memory." It remembers what happened minutes or hours ago to help clean up the current noise. It's like a detective who remembers a suspect's behavior from yesterday to solve a crime today.

How They Trained the AI (The "Fake Data" Trick)

Real flight data with perfect "clean" magnetic signals is very rare and hard to get. To teach the AI, the team had to be creative.

The Recipe: They used a supercomputer model of the Earth's magnetic field (the World Magnetic Model) and mixed it with a "Generative Adversarial Network" (a type of AI that creates fake data).
The Result: They generated thousands of hours of realistic, synthetic flight data. It's like training a pilot in a flight simulator that is so realistic, the pilot learns to handle the turbulence before ever leaving the ground.

The Results: Does it Work?

The team tested their new "Physics-Aware" AI against old methods and standard AI.

The Winner: The AI that followed the physics rules (Divergence-Free + E(3)-Equivariance) and used the "Contiformer" architecture was the clear winner.
The Improvement: It reduced errors significantly (by about 30% to 57% compared to other methods).
The Takeaway: By forcing the AI to obey the laws of physics, it became much better at separating the "signal" (the Earth's map) from the "noise" (the plane's interference).

Summary

The paper presents a new way to clean up magnetic sensor data for planes. Instead of just letting an AI guess, they built the AI with hard-coded rules of physics (magnetic loops and rotation rules) and gave it a long-term memory to handle slow-moving noise. This makes the AI much better at finding the Earth's magnetic "fingerprint" even when the plane is making a lot of magnetic noise.

Technical Summary: Physics-Aware Neural Networks for Denoising in Magnetic Navigation

Problem Statement
Magnetic-anomaly navigation offers a critical alternative to Global Navigation Satellite Systems (GNSS) in environments where signals are jammed or spoofed. However, airborne magnetic navigation faces a fundamental challenge: the aircraft itself generates magnetic noise from its chassis, engines, and instruments, which corrupts the local geomagnetic field measurements required for geolocation. While classical methods like the Tolles-Lawson model address deterministic interference, they fail to adequately handle the stochastic noise inherent in operational data. Furthermore, existing deep learning approaches often lack the physical consistency required to model magnetic fields accurately, leading to error-prone navigation.

Methodology
The authors propose a first-principles approach that integrates physical laws directly into neural network architectures to denoise magnetometer data. The methodology is built upon three core pillars:

Physics-Aware Constraints:
- Divergence-Free Constraint: To satisfy Maxwell's second law ( $\nabla \cdot \mathbf{B} = 0$ ), the network does not output the magnetic field $\mathbf{B}$ directly. Instead, it outputs a vector potential $\mathbf{A}_\theta$ . The magnetic field is constructed as the curl of this potential ( $\mathbf{B} = \nabla \times \mathbf{A}_\theta$ ). This guarantees the divergence-free property identically, as $\nabla \cdot (\nabla \times \mathbf{A}) = 0$ for any smooth vector field.
- E(3)-Equivariance: To ensure predictions remain consistent regardless of the sensor's position and orientation, the network enforces E(3)-equivariance (invariance under rigid transformations). This is implemented using geometric tensors represented by spherical harmonics. The architecture utilizes tensor products and Clebsch-Gordan coefficients to ensure that if the input is rotated, the output rotates correspondingly.
Architectural Design for Temporal Dynamics:
Recognizing that magnetometer data is irregularly sampled and exhibits long-term dependencies (e.g., eddy currents, thermal drift), the authors model the system using continuous-time dynamics.
- Neural ODEs: The hidden state $z(t)$ evolves via a neural ordinary differential equation ( $\dot{z}(t) = f_\theta(z(t), t)$ ), allowing the model to handle arbitrary time intervals $\Delta t_i$ .
- Long-Term Memory: The architecture incorporates mechanisms to retain long-range context, essential for modeling noise sources that decay slowly over time.
- Evaluated Architectures: The study evaluates four backbone architectures: Multi-Layer Perceptrons (MLP), 1D Convolutional Neural Networks (CNN), Liquid Time-Constant (LTC) networks, and Contiformers. The Contiformer, which inherently possesses continuous-time dynamics and long-term memory, is highlighted as the most effective.
Synthetic Data Generation:
Addressing the scarcity of real-world flight data, the authors developed a conditional Generative Adversarial Network (cGAN). This model learns the joint distribution of windowed sensor readings conditioned on discrete context labels (e.g., flight ID, attitude). The generator uses a 2-layer LSTM to produce realistic, temporally consistent magnetic field sequences based on the World Magnetic Model (WMM).

Key Contributions

Novel Constraints: Introduction of divergence-free vector fields and E(3)-equivariance as hard constraints within neural networks for magnetic navigation, ensuring adherence to Maxwell's equations and rotational consistency.
Architectural Identification: Identification of continuous-time dynamics and long-term memory as critical properties for modeling magnetic time-series data, with the Contiformer architecture demonstrating superior performance.
Data Synthesis: Development of a synthetic dataset generation pipeline using cGANs and the WMM to overcome data scarcity.
System Implementation: Design and evaluation of a complete denoising pipeline that outperforms both classical compensation models and unconstrained deep learning approaches.

Results
The system was evaluated on real flight data (MagNav dataset) and synthetic data generated by the cGAN. Performance was measured using Root Mean Squared Error (RMSE) and Signal-to-Noise Ratio (SNR).

Impact of Constraints: Adding physics-aware constraints consistently improved performance across all architectures. The combination of both constraints yielded the highest gains. For instance, in the MLP architecture, the test RMSE dropped by 57% when both constraints were applied.
Architecture Comparison: The Contiformer architecture significantly outperformed MLPs, CNNs, and LTCs. The best-performing model (Contiformer with both constraints) achieved a test RMSE of 13.09 nT and an SNR of 56.47 dB, compared to the baseline MLP test RMSE of 70.47 nT.
Ablation Studies: The study confirmed that while individual constraints (divergence-free or equivariance) improved results, their combination provided the most robust denoising, effectively acting as an implicit regularizer.

Significance
The paper claims that embedding fundamental physical principles directly into network architectures provides a generalizable framework for spatio-temporal systems. By constraining the space of admissible field functions to physically plausible solutions, the model improves predictive accuracy and physical plausibility without relying solely on data volume. The work demonstrates that physics-aware constraints can effectively bridge the gap between classical physical models and modern deep learning, offering a robust solution for magnetic anomaly navigation in data-scarce and noisy environments. The authors note that this approach is applicable beyond magnetic navigation to other dynamical systems governed by physical laws.