Neural Fields for NV-Center Inverse Sensing

This paper introduces NeTMY, an amortization-free coordinate neural field framework that overcomes center-collapse failures in NV-center magnetic noise sensing by replacing scalar forward approximations with a corrected tensor operator and employing specialized optimization strategies to achieve superior sparse source localization.

The Gist

The Big Picture: Solving a "Blind" Puzzle

Imagine you are trying to figure out where a group of people are standing in a dark room. You can't see them, but you have a microphone that picks up the sound of their footsteps. However, the microphone is weird:

1. It distorts the sound: The sound gets quieter the further the person is from the mic.
2. It mixes sounds: If two people are close together, their footsteps blend into one noise.
3. It's noisy: There is static in the recording.

Your goal is to look at the messy audio recording and draw a map showing exactly where each person is standing. In the scientific world, this is called an inverse problem: working backward from a messy result to find the original cause.

The paper focuses on a specific type of "microphone" called a Nitrogen-Vacancy (NV) center (a tiny defect in a diamond) that senses magnetic "noise" from tiny spinning particles (spins) in a material.

The Problem: The "Bad Map" vs. The "Good Map"

The researchers found that most scientists use a simplified, "lazy" way to model how the microphone works. They call this the Scalar Approximation.

• The Analogy: Imagine trying to figure out where people are by squaring the volume of the sound. If two people are talking, you just add their volumes and square the result.
• The Flaw: This creates "ghosts." Mathematically, this method invents fake connections between people who aren't actually interacting. When you try to solve the puzzle using this bad map, the computer gets confused and thinks everyone is standing right in the center of the room, even if they are scattered around the edges. The researchers call this "Center-Collapse."

The paper introduces a Tensor Power-Summed Operator.

• The Analogy: This is the "physics-accurate" map. Instead of squaring the total volume, it calculates the energy of each person's footsteps separately and then adds them up. It respects the fact that the people are independent.
• The Result: This map doesn't have the "ghost" connections. It reveals that the "Center-Collapse" was an illusion caused by the bad math. When you use the good map, the puzzle becomes much harder to solve because the clues are more subtle, but the answer is physically real.

The Solution: NeTMY (The Smart Detective)

The researchers built a new tool called NeTMY to solve this puzzle. Instead of using a pre-trained AI (which learns by looking at thousands of examples) or a simple math formula, NeTMY acts like a detective who solves the case from scratch every time.

Here is how NeTMY works, using three key tricks:

1. The "Zoom-Out to Zoom-In" Strategy (Multiscale Optimization)

• The Problem: If you try to find a tiny speck of dust in a photo by looking at every pixel at once, you get overwhelmed by noise.
• The Trick: NeTMY starts by looking at a blurry, low-resolution version of the map. It finds the general shape of the crowd first. Once it knows where the crowd is roughly located, it zooms in to find the exact spots of the individuals. This prevents the detective from getting lost in the static.

2. The "Smoothie" Filter (Neural Field Parameterization)

• The Problem: When the "bad math" (Center-Collapse) happens, the computer tries to move everything to the center in one giant, jerky leap.
• The Trick: NeTMY doesn't move pixels directly. Instead, it moves a "smoothie" (a continuous mathematical curve) that represents the map. If the computer wants to move a pixel, it has to move the whole smooth curve. This acts like a filter that smooths out the jerky, center-pulling forces. It forces the solution to be physically reasonable, preventing the "Center-Collapse" failure.

3. The "Annealing" Schedule (Turning Up the Volume)

• The Problem: The high-frequency details (the tiny, sharp edges of the spins) are very hard to hear over the noise.
• The Trick: NeTMY starts by only listening to the low, rumbling sounds (the big shapes). As it gets better, it slowly "turns up the volume" on the high-pitched, sharp sounds. This lets it build a solid foundation before trying to hear the tiny details.

The Results: Who Won the Puzzle?

The researchers tested NeTMY against old-school math methods (like Tikhonov and ADMM) and other AI methods.

• The Old Methods: When using the "physics-accurate" map, these methods failed miserably. They all fell into the "Center-Collapse" trap, drawing a big blob in the middle of the room, missing the actual people scattered around.
• The Supervised AI: Methods that learned from training data failed because they were trained on "crowded" scenes but tested on "sparse" (few people) scenes. They couldn't generalize.
• NeTMY: It won. It successfully reconstructed the scattered, sparse sources without collapsing them into the center. It found the right locations and the right shapes better than anyone else.

Why This Matters (According to the Paper)

The paper argues that this isn't just about diamond sensors. It proves that how you model the physics matters more than you think.

• If you use a simplified model, your AI might learn to cheat and find fake solutions (like the center collapse).
• If you use a faithful, complex model, the problem becomes harder, but you need a smarter solver (like NeTMY) to handle it.

The authors conclude that NV sensing is a perfect "testbed" (a practice arena) for testing these physics-faithful AI methods because the physics is so delicate and the "bad math" traps are so obvious.

In short: They fixed the "map" (the physics model) so it didn't lie, and they built a new "detective" (NeTMY) that is smart enough to solve the puzzle without getting tricked by the noise or collapsing into the center.

Technical Summary

Technical Summary: Neural Fields for NV-Center Inverse Sensing

Problem Formulation

The paper addresses the inverse problem of reconstructing sparse, fluctuating spin source distributions and their local Larmor response from noisy magnetic-noise spectra measured by Nitrogen-Vacancy (NV) centers in diamond. Unlike standard magnetic field imaging which recovers static fields, this task involves inferring a sparse density field $\rho$ and a spectral field $\omega_L$ from frequency-dependent noise spectra $S_{obs}$.

The problem is characterized by severe ill-posedness due to four structural pathologies:

1. Exponential Frequency Suppression: High-spatial-frequency features are exponentially suppressed by the dipolar Green's tensor decay ($e^{-kz_0}$), making high-frequency recovery unstable.
2. Finite-Window Center Bias: The convolutional footprint of a source is more fully observable at the center of the sensing window than at the boundaries, creating a gradient bias toward the center even from uniform initialization.
3. Max-Normalization Peak Coupling: The standard normalization of noise spectra introduces a non-local gradient term concentrated at the current peak pixel, which self-reinforces any incipient peak, particularly at the window center.
4. Resolution-Limited Merging: Sources separated by less than the effective point-spread width ($\sim z_0$) cannot be disambiguated, and the Larmor frequency is only identifiable on the support of the density.

A critical finding of the work is that the choice of the forward operator significantly alters the optimization landscape. The authors compare a simplified scalar/coherent operator ($F_1$), which squares a coherently summed field, against a tensor power-summed operator ($F_2$), which sums per-channel noise powers. While $F_1$ is computationally cheaper, $F_2$ is physically more faithful for incoherent thermal fluctuators. The paper demonstrates that $F_1$ masks a "center-collapse" failure mode where free-density optimizers converge to a central artifact, whereas $F_2$ exposes this pathology, making the inverse problem significantly harder for standard solvers.

Methodology: NeTMY

The authors propose NeTMY (Neural Tensor Magnetic Yield), an amortization-free coordinate neural field solver. Unlike supervised methods that require paired training data (which are scarce for NV sensing) or classical methods that optimize the density field directly, NeTMY represents the unknown density $\rho$ and Larmor field $\omega_L$ as the output of a coordinate Multi-Layer Perceptron (MLP). The network parameters are optimized per measurement instance against a single observed spectrum.

Key design components include:

• Coordinate Neural Field with Annealed Positional Encoding: The MLP takes spatial coordinates as input, augmented with Fourier features. These features are "annealed" (gradually activated) during training, allowing the network to fit low-frequency structures before high-frequency details, addressing the exponential frequency suppression.
• Gated Density and Larmor Heads: The density output uses a gated softplus to enforce non-negativity and allow the network to drive regions to near-zero without saturation. The Larmor output is masked by the predicted density support, ensuring gradients only flow where the data constrains the solution.
• Multiscale Curriculum: Optimization proceeds in two stages, starting with a coarse grid resolution to recover the global support and refining on a finer grid to capture high-frequency details.
• Physics-Faithful Losses: The objective function combines a canonical log-MSE data fidelity term with specific physics-driven losses: a mean-normalized noise-map loss to anchor gradients on the support, and a direct-density loss to proxy the amplitude.
• Energy-Anchored Scale Correction: Since max-normalization renders the absolute density scale unidentifiable, a post-processing step rescales the predicted density based on the ratio of observed to predicted total energy.

Key Contributions

1. Physics-Faithful Forward Operator: The authors formulate NV noise sensing as a differentiable inverse problem using a tensor power-summed operator ($F_2$) that avoids nonphysical cross-terms present in simplified scalar solvers ($F_1$). They show that this choice fundamentally reshapes the inverse landscape.
2. NeTMY Solver: They introduce an amortization-free coordinate neural field solver that reconstructs sparse fields without paired density labels. The method leverages the parameterization geometry to smooth updates and mitigate center-collapse.
3. Mechanistic Analysis of Optimization Geometry: The paper provides a theoretical and empirical explanation for why NeTMY succeeds where classical solvers fail. It demonstrates that free-density solvers execute the raw, biased density-space gradient, while NeTMY's parameterization acts as a positive semi-definite filter ($G_\theta = J_\theta J_\theta^\top$) that redistributes the gradient, preventing the singular center spike that leads to collapse.
4. Benchmarking and Real-Data Validation: The authors establish a cross-fidelity benchmark where data is generated by a highly accurate direct simulator ($F_3$) but inverted using $F_1$ or $F_2$. They further validate the operator-fidelity gap on a real $\alpha$-RuCl$_3$ dataset, showing that the more faithful operator ($F_2$) yields better consistency with physical priors (depth and amplitude) and a better-conditioned loss landscape.

Results

• Performance on Synthetic Benchmarks: On a cross-fidelity benchmark (512 samples generated by $F_3$, inverted by $F_1$ or $F_2$), NeTMY achieves the best localization (Hungarian F1) and distributional (Sliced Wasserstein Distance) metrics, particularly under the physically correct $F_2$ operator. Classical methods like Tikhonov and ADMM suffer from center-collapse under $F_2$, resulting in poor localization.
• Mechanism Verification: Experiments confirm that free-density solvers exhibit a strong center-bias in the initial gradient (center-to-outer ratio of ~18x) and get trapped in a local minimum separated from the ground truth by an energy barrier. NeTMY's first-step update is spatially distributed (center-to-outer ratio ~1.6x), avoiding this trap.
• Real-Data Consistency: On the $\alpha$-RuCl$_3$ dataset, the $F_2$ operator allows for a physically consistent depth-amplitude calibration (recovering the expected depth for 1/8 of NVs within the prior range, whereas $F_1$ fails for all). Furthermore, the loss landscape under $F_2$ is a well-conditioned parabolic bowl, while $F_1$ results in a degenerate valley, confirming the superior identifiability of the faithful operator.
• Ablation: Removing components like annealed positional encoding, multiscale scheduling, or the gating mechanism significantly degrades performance, confirming that the design choices directly address the identified ill-posedness pathologies.

Significance and Claims

The paper positions NV quantum sensing as a rigorous testbed for physics-faithful neural inverse problems. It argues that the fidelity of the forward operator is not merely a matter of measurement accuracy but fundamentally alters the geometry of the inverse problem, creating failure modes (like center-collapse) that are hidden under simplified approximations.

The authors claim that NeTMY's success stems from the interplay between representation geometry and optimization, rather than just expressivity. By using a coordinate neural field, the method implicitly filters the raw gradient, smoothing out pathological biases inherent in the physical forward model. The work suggests that for scientific sensing tasks where paired labels are unavailable and physical models are complex, amortization-free neural fields coupled with faithful forward operators offer a robust alternative to both classical regularized solvers and supervised deep learning.

The paper remains modest regarding its scope, acknowledging that NeTMY is computationally slower than classical baselines (roughly 100x) and is currently limited to the fluctuation-dominated dipolar regime. It does not claim to solve all quantum sensing modalities but establishes a framework for addressing operator-fidelity-induced optimization challenges in sparse reconstruction.