Physics-Informed Deep Image Prior Reconstruction of In-Plane Magnetization from Scanning NV Magnetometry

This paper demonstrates a physics-informed deep image prior (DIP) framework that reconstructs complex in-plane magnetization patterns from scanning NV magnetometry data without requiring pre-trained datasets, showing that using optimally aligned spatial masks significantly improves reconstruction accuracy and signal-to-noise ratio.

The Gist

Imagine you are trying to figure out what a complex, beautiful stained-glass window looks like, but you aren't allowed to look at it directly. Instead, you are standing in a dark room, and the only thing you can see is the pattern of light and shadows cast on the floor by a single lamp.

This is the fundamental problem the scientists in this paper are solving.

The Problem: The "Shadow" Mystery

In the world of tiny electronics (spintronics), scientists need to see how magnetism is arranged inside microscopic materials. They use a tool called NV Magnetometry, which is like a super-sensitive "light sensor" that detects the magnetic "shadows" (called stray fields) cast by these materials.

The problem is that the "shadow" doesn't tell the whole story. Because of the laws of physics, many different magnetic patterns can cast the exact same shadow. If you only look at the shadow, you might guess the wrong pattern. This is what scientists call an "ill-posed problem"—it’s like trying to guess the shape of a sculpture just by looking at its silhouette.

The Solution: The "Smart Artist" (DIP)

To solve this, the researchers used a specialized AI called a Deep Image Prior (DIP).

Think of this AI not as a student who has memorized a textbook (which would be "supervised learning"), but as a highly skilled artist with a strong sense of logic.

Instead of showing the AI millions of pictures of magnets to study, they give the AI a blank canvas and a set of "rules of the universe" (the physics). The AI then starts doodling random shapes. Every time it draws a shape, it asks itself: "If this were the real magnet, would it cast the shadow I actually see on the floor?"

If the answer is "No," the AI erases and tries again. Because the AI is built with a specific structure (a "convolutional autoencoder"), it has a natural "artistic style" that prefers smooth, realistic shapes over messy, random noise. This helps it "ignore" the static and noise in the data and focus on the real magnetic patterns.

The Secret Ingredient: The "Stencil"

Even with a smart artist, there is still too much guesswork. To help the AI, the scientists provide a "mask"—think of this as a stencil.

If they know the magnet is shaped like an ellipse, they give the AI a stencil of an ellipse. This tells the AI: "Don't bother drawing magnetism in the empty space around the object; only focus inside this shape."

The researchers discovered something fascinating: The stencil is a diagnostic tool. If they use a stencil that is slightly tilted or the wrong shape, the AI struggles to "draw" the right answer. But when the stencil aligns perfectly with the real object, the AI finds the answer incredibly fast and accurately. It’s like trying to trace a drawing; if your paper is crooked, you’ll never get it right!

Why Does This Matter?

As we try to make computers faster and more energy-efficient (using "spintronics"), we need to see exactly how magnetism behaves at a microscopic level.

This paper proves that we don't need massive supercomputers or giant databases of images to "see" the invisible. By combining smart physics rules with a clever AI artist, we can reconstruct the hidden magnetic maps of the future with incredible precision.

Technical Summary

Technical Summary: Physics-Informed Deep Image Prior Reconstruction of In-Plane Magnetization from Scanning NV Magnetometry

1. Problem Statement

The reconstruction of magnetization ($\mathbf{M}$) from stray magnetic field measurements ($\mathbf{B}$) is a fundamental challenge in spintronics and nanoscale magnetic imaging. This task is inherently ill-posed because the magnetostatic transformation is non-unique; multiple different magnetization configurations can produce identical stray field distributions.

While Scanning Nitrogen-Vacancy (NV) magnetometry provides high spatial resolution (down to 20 nm) and high sensitivity, extracting the underlying in-plane magnetization from the measured field projection ($B_{NV}$) remains difficult. Existing deep learning solutions often require massive, pre-trained datasets (supervised learning), which are difficult to obtain for specific, novel magnetic nanostructures.

2. Methodology

The authors propose a Physics-Informed Deep Image Prior (DIP) framework. Unlike supervised learning, this approach does not require a training dataset; instead, it learns to reconstruct a single specific instance by leveraging the architectural properties of a neural network as a regularizer.

• Architecture: The researchers utilize a Convolutional Autoencoder (CAE) consisting of an encoder (downsampling via max pooling and convolutional layers) and a decoder (upsampling via bilinear interpolation). Notably, they intentionally omitted "skip connections" (common in U-Net architectures) to provide stronger implicit regularization, which helps prevent the network from overfitting to measurement noise.
• Physics-Informed Loss Function: The network is trained by minimizing the Mean Square Error (MSE) between the measured NV signal ($B_{NV, \text{meas}}$) and a predicted signal ($B_{NV, \text{model}}$). The predicted signal is generated by:
  1. Taking the network's output (the reconstructed $\mathbf{M}$).
  2. Applying a magnetostatic forward model in Fourier space to calculate the resulting stray field.
  3. Projecting that field onto the NV quantization axis.
• Physical Constraints: To resolve the non-uniqueness of the inverse problem, the authors implement two critical constraints:
  • Spatial Masking: A user-defined mask (0 to 1) is applied to the output to ensure magnetization only exists within the physical boundaries of the nanostructure.
  • Solution Initialization: The model uses an assumed magnetization distribution to guide convergence.
• Activation Function: The authors found that using a Tanh activation function for the final layer provided superior convergence and better signal-to-noise ratios (SNR) compared to a Sigmoid function.

3. Key Contributions

• Resource-Efficient Reconstruction: Demonstrated that a simplified CAE architecture can achieve high-quality reconstruction without the need for large, pre-trained datasets.
• Diagnostic Masking Tool: Discovered that the orientation of the user-defined spatial mask significantly impacts convergence. The alignment of the mask with the actual domain structure improves the reconstruction SNR by up to 3 dB.
• Experimental Validation: Provided the first experimental validation of DIP-based reconstruction for Landau and dipole domain patterns in Permalloy nanostructures using scanning NV magnetometry, corroborated by Magnetic Force Microscopy (MFM) and micromagnetic simulations (mumax3).

4. Results

• Landau Domain Reconstruction: For a complex Landau pattern, the model achieved an optimal alignment (0° rotation) with an RMSE of 0.56 mT and an SNR of 4.94 dB.
• Dipole Domain Reconstruction: For a simpler dipole configuration, the model achieved an RMSE of 0.34 mT when the mask was correctly oriented.
• Sensitivity Analysis:
  • Mask Orientation: The error (MSE) was lowest when the mask matched the true domain geometry, proving that the convergence rate can serve as a diagnostic tool to identify unknown magnetic textures.
  • Standoff Distance: The reconstruction accuracy is highly sensitive to the distance between the NV tip and the sample. There is a non-monotonic relationship where too large a distance filters out high-frequency details, while too small a distance leads to signal attenuation and noise dominance.

5. Significance

This work provides a practical, computationally efficient framework for the quantitative analysis of magnetic nanostructures. By combining the implicit regularization of deep learning with the rigorous constraints of magnetostatic physics, the authors have bypassed the need for massive training data. This methodology is particularly relevant for the development of next-generation spintronic technologies, such as racetrack memory, where characterizing complex, in-plane magnetic textures at the nanoscale is essential.