Skip to content

High-resolution wide-field magnetic imaging with sparse sampling using nitrogen-vacancy centers

FieldValue
JournalOpen MIND
Published2026-01-31
AuthorsKeqing Liu, Jiazhao Tian, Bokun Duan, Hao Zhang, Kangze Li
DOI10.48550/arxiv.2602.00679

Twenty-five simulated reconstruction samples produced a 10,000-pixel magnetic-field map with a structural similarity index above 0.999 for smooth fields containing one to three extrema. Mean-adjusted Bayesian estimation reduced mean absolute error and root mean square error from the (10^{-2}) level to the (10^{-4}) level. The evidence is entirely simulated, with no experimental acquisition time or calibrated spatial resolution reported.

Conventional wide-field NV microscopy reads many NV ensembles in parallel, so optical diffraction and sensor-to-sample standoff set lateral resolution. Single-NV probes can reach genuine nanoscale resolution, but each pixel typically requires a millisecond or longer dwell. Their fields of view are limited to a few to a few tens of square micrometres, and acquisition takes minutes or more.

Regular wide-field microscopes cover fields from tens of micrometres to the millimetre scale, but effective resolution usually remains in the few-hundred-nanometre range. Super-resolution methods can reach about 10 to 100 nm, yet most require scanning or repeated patterned excitation. The paper therefore targets acquisition sparsity rather than a new optical resolution mechanism.

MABE reconstructs the full distribution from a correlated error model and a small set of measurements. The reported 25-point result applies to representative smooth fields, not arbitrary current patterns.

The simulation models negatively charged NV centers with an electron spin (S=1) and zero-field splitting (D=2\pi\times2.88) GHz. Microwave (\pi) pulses at each zero crossing of the AC field accumulate a spin phase. Ramsey readout then gives (P_0=\frac{1}{2}[1+\cos\Phi(t)]), while the field amplitude follows (B=g_{\mathrm{AC}}/\gamma=\pi\omega/(4\gamma)).

Static inhomogeneous detuning had a FWHM of (2\pi\times26.5) MHz. Dynamic noise followed an Ornstein-Uhlenbeck process with a 20 µs correlation time and a stated diffusion scale of (2\pi\times50) kHz. Each XY-8 sequence used 50 ns pulses, 950 ns intervals, and 400 µs total evolution for (N=50). Population contrast fell below 0.2 under these conditions.

Phase-modulated pulses improve sensitivity

Section titled “Phase-modulated pulses improve sensitivity”

Phase-modulated microwave pulses were optimized against the noise model to raise gate fidelity and preserve spin contrast. The controls used (a=0.0628), (b=0.0830), and (\nu=0.0316). Optimization sampled 15 detuning values across ([-W,W]), covering about 98% of the static distribution. Drive amplitude was constrained below (2\pi\times10) MHz.

Pauli-X and Pauli-Y average gate fidelity rose from about 0.43 to 0.68. The optimized XY-8 sequence reached a noisy population contrast of about 0.6, compared with below 0.2 for rectangular pulses. The readout-noise standard deviation was set to (\sigma=10^{-2}) for the sensitivity calculation.

At (N=16) and (t=128) µs, sensitivity improved from 0.94 to 0.45 nT/√Hz. At (N=32) and (t=256) µs, it improved from 1.33 to 0.64 nT/√Hz. The population slope increased by more than twofold, giving the paper’s approximate twofold sensitivity gain.

MABE begins with a correlated error model for the magnetic field. Its correlation decays as (\exp[-d]), and the distance function uses nonnegative weights (\alpha_h) and exponents (P_h) between 1 and 2. The model parameters are optimized by maximizing the likelihood of the sampled measurements.

A best linear unbiased predictor then estimates the field at each output pixel from the sample correlations. Sampling locations began as a uniform grid with uniform displacements bounded by 30% of the field-of-view size. For each test field, 25 locations generated a 100 × 100 pixel reconstruction. Direct point interpolation from the same samples captured only coarse features.

Ten reference points were measured in addition to the 25 reconstruction samples. Their nominal values spanned the sampled value range, and they used the same protocol under identical noise conditions. Proportional calibration produced smaller absolute errors and a narrower error distribution than additive bias calibration. The adjusted measurements, rather than the raw values, entered the final MABE reconstruction.

The study assessed mean absolute error, root mean square error, peak signal-to-noise ratio, coefficient of determination, and structural similarity. MABE reduced mean absolute error and root mean square error from the (10^{-2}) level to the (10^{-4}) level for all three distributions.

Field distributionPSNRR²SSIM
Single extremum30.58 dB0.9830.9998
Two extrema30.21 dB0.9880.9998
Three extrema28.3 dB0.9780.9998

Direct outputs from 25 measurements had a peak signal-to-noise ratio of about 2.3 dB. Their coefficients of determination were negative, and their structural similarity was about 0.13. The MABE metrics therefore improved consistently across error size, signal level, fit quality, and structure. These values apply to the simulated smooth fields and the implemented measurement noise.

The sampling-density test used 25, 36, 64, 81, and 100 points. Each point on the mean absolute error curve averaged 20 independent reconstructions with different sample sets. For the selected three-extrema field, 25 points held mean absolute error near the (10^{-4}) level. The authors selected 25 points as a practical balance between acquisition time and reconstruction fidelity.

Five patterns were compared: random, spiral, square-loop, serpentine, and grid. Grid sampling produced the lowest mean absolute error and became the default pattern. Its starting coordinates came from uniform grid centers with bounded perturbations.

At 100 sampling points, structural similarity exceeded 0.9999 for all three fields. Raising the count from 25 to 100 quadrupled acquisition points but produced only modest metric improvements. Boundary agreement improved gradually as sampling density increased.

The main limitation is that all results come from simulations of fields with one to three extrema. The authors state that high fidelity at low sampling density depends on relatively smooth, structurally simple distributions. More complex spatial patterns may require higher sampling densities or stronger inference assumptions. Compressed sensing, Gaussian-process adaptive sampling, and deep image priors are proposed as possible directions, but they were not tested here.

The method is training-free and structurally simpler than supervised reconstruction methods. The authors also suggest extending it to point-scanning NV sensors, SQUIDs, Hall probes, and magnetic tunnel junctions. Those extensions remain suggestions rather than measured demonstrations.

A (10^4)-pixel output does not establish a physical resolution because the paper reports no calibrated pixel size. It also reports no experimental wall-clock imaging time. The nominal reference values required for calibration are not defined for a practical instrument.

For device engineering, sparse inference may reduce acquisition count when the field is smooth and the correlation model is valid. It does not remove the need for calibrated references or experimental validation. Localized currents can create more complex spatial patterns, so sampling density should be tested against the field structure expected from the target device. Device design must still account for optical diffraction and sensor-to-sample standoff, which the paper identifies as joint limits on lateral resolution.

The paper’s magnetometer uses a diamond sensing plate with near-surface NV ensembles for wide-field magnetic imaging. 6CCVD’s diamond materials catalog lists MPCVD-grown single-crystal and polycrystalline plates and wafers, with single-crystal polishing to Ra below 1 nm. NV formation and near-surface layer depth must be confirmed through the engineering support page before selecting the sensing plate.

View Original Abstract

Nitrogen-vacancy (NV) centers in diamond enable quantitative magnetic imaging, yet practical implementations must balance spatial resolution against acquisition time (and thus per-pixel sensitivity). Single-NV scanning magnetometry achieves genuine nanoscale resolution, nonetheless requires typically a slow pixel-by-pixel acquisition. Meanwhile, wide-field NV-ensemble microscopy provides parallel readout over a large field of view, however is jointly limited by the optical diffraction limit and the sensor-sample standoff. Here, we present a sparse-sampling strategy for reconstructing high-resolution wide-field images from only a small number of measurements. Using simulated NV-ensemble detection of ac magnetic fields, we show that a mean-adjusted Bayesian estimation (MABE) framework can reconstruct 10000-pixel images from only 25 sampling points, achieving SSIM values exceeding 0.999 for representative smooth field distributions, while optimized dynamical-decoupling pulse sequences yield an approximately twofold improvement in magnetic-field sensitivity. The method further clarifies how sampling patterns and sampling density affect reconstruction accuracy and suggests a route toward faster and more scalable magnetic-imaging architectures that may extend to point-scanning NV sensors and other magnetometry platforms, such as SQUIDs, Hall probes, and magnetic tunnel junctions.