Skip to content

Accurate magnetic field estimation in fiber-coupled nanodiamond NV ensembles using model-assisted regression

FieldValue
JournalOptics Express
Published2026-02-27
AuthorsJian Gao, Doudou Zheng, chenchen tang, Xiaoqi Gao, jiayuan fu
DOI10.1364/oe.589141
Citations1

Model-assisted regression reduced the magnetic-field RMSE from 37.21 µT to 8.99 µT under identical acquisition conditions at an optimized intrinsic field of 640 µT. It also reduced relative slope error from 0.60059 to 0.0386, a reported 15.6-fold improvement over Lorentzian fitting.

The ensemble response differs from a single NV response

Section titled “The ensemble response differs from a single NV response”

Randomly oriented NV centers produced overlapping ODMR resonances rather than one clean Lorentzian peak. Each axis projected the applied field differently, so orientation averaging broadened and asymmetrically distorted the ensemble spectrum. This broke the assumed frequency-to-field linearity and biased Lorentzian fitting, especially for larger perturbations.

The fiber-nanodiamond model includes both NV orientation statistics and the fiber numerical aperture in the fluorescence collection weight. It defines K as the slope of the ensemble-averaged resonance frequency against a small perturbation around B0. This coefficient provides the physical scale for converting frequency shifts to fields in the ensemble.

Simulations showed that increasing B0 initially increased K before saturation from orientation averaging. A larger fiber numerical aperture collected more large-angle axes with weaker Zeeman projections, reducing or saturating K. The model depends on averaging over many NV centers and is most suitable for dense nanodiamond ensembles.

Fisher information selects the operating field

Section titled “Fisher information selects the operating field”

Fisher information was used to select the intrinsic field before training the regression model. The analysis treated noise as frequency independent and dominated by photon shot noise and fluorescence fluctuations under continuous-wave ODMR. Because the noise level acts as a constant prefactor, it does not shift the predicted optimum.

The model numerically evaluated K for combinations of fiber numerical aperture and bias field B0. Fisher information was then calculated from the modeled ensemble ODMR spectrum within the Cramér-Rao framework for minimum field uncertainty.

The information maximum occurred at approximately 640 µT. The measured mean absolute error also reached its minimum at that operating point. Values away from 640 µT had lower Fisher information and higher reconstruction error.

The authors calibrated K independently by applying small, known perturbations around 640 µT and fitting the ensemble-averaged resonance shifts. This calibration gave the regression a physical frequency-to-field scale. Intrinsic-field-guided Gaussian regression, or IFGGR, then corrected residual distortion from orientation averaging, numerical-aperture effects, and experimental fluctuations.

Acquisition provides calibrated frequency-shift labels

Section titled “Acquisition provides calibrated frequency-shift labels”

A 532 nm laser was coupled into a multimode fiber, which also collected the emitted fluorescence. A horn antenna delivered a swept microwave signal from 2.8 to 3.1 GHz. A calibrated Helmholtz coil pair generated known perturbations around a fixed intrinsic field B0.

All measurements were conducted at room temperature. Each field setting used ten repeated ODMR acquisitions to suppress shot noise and stabilize sampling. Each normalized spectrum received a frequency-shift label ∆f derived from the known perturbation and calibrated ensemble response.

The coil calibration fixed the reference field, while the model mapped that known perturbation to ∆f. The labels therefore supported independent evaluation of spectral extraction without relying on line-position readings from each distorted spectrum.

The dataset was shuffled and divided into training and test subsets at a 4:1 ratio. During unknown-field reconstruction, IFGGR inferred ∆f from the distorted spectrum. Each point in the later error sequence came from one ODMR acquisition with a 30 s integration time.

The regression model corrects spectral distortion

Section titled “The regression model corrects spectral distortion”

Nine Gaussian process regression configurations were compared using RMSE, R², and MAE. The weighted score assigned 0.2 to RMSE, 0.6 to R², and 0.2 to MAE. The best configuration used a constant mean function, a rational quadratic covariance kernel, and a Gaussian likelihood.

The rational quadratic kernel was chosen for the multi-scale variations caused by random NV orientations, strain inhomogeneity, overlapping resonances, and measurement noise. Against the squared exponential kernel under the same preprocessing and split, it lowered MAE from 17.82 µT to 15.16 µT. It also lowered RMSE from 18.75 µT to 15.77 µT, an improvement of approximately 15% in each metric.

Both kernels used identical preprocessing, the same training-validation split, and comparable hyperparameter optimization budgets. The rational quadratic kernel also produced lower field-resolved errors across most of the tested magnetic-field range.

Tsukamoto et al. used Gaussian process regression to infer field magnitude directly from ensemble spectra. Other studies tested neural networks for spectrum-to-field inference and embedded sensing. Those end-to-end mappings often made calibration and operating-point selection implicit. The method examined here instead fixes B0 and K through physical analysis, then uses regression for residual frequency-shift extraction.

Table 1 compares model-assisted regression with Lorentzian fitting under identical acquisition conditions.

MethodMAE, µTRMSE, µTBias, µTMaximum absolute error, µT
Model-assisted regression7.7058.99477.696316.135
Lorentzian36.67637.207-36.67648.188

Model-assisted regression reduced RMSE by about a factor of four and cut the systematic deviation. The same ODMR spectra were processed by both methods, so the comparison isolated estimator behavior rather than acquisition differences.

The positive 7.6963 µT bias indicated a small mean overestimate. Lorentzian fitting instead showed a 36.676 µT mean underestimate. Relative slope error fell from 0.60059 for Lorentzian fitting to 0.0386 for IFGGR. That corresponded to the reported 93.57% reduction and 15.6-fold improvement in relative slope error.

IFGGR learned the inverse mapping without assuming a fixed parametric line shape. The independently calibrated K preserved the physical field scale. These results apply at B0 = 640 µT and a 30 s integration per spectrum.

Sensitivity follows the corrected error floor

Section titled “Sensitivity follows the corrected error floor”

At 640 µT, the measured ODMR linewidth was 28 MHz, contrast was 0.07, and detected fluorescence power was 20 µW. Using a Lorentzian line-shape correction factor PL of 0.77 gave a calculated photon shot-noise sensitivity of 1.37 nT/√Hz.

The shot-noise reference used the measured optical and spectral parameters rather than the regression residual itself. Absolute sensitivity was calculated from the standard deviation of block-averaged errors after removing the mean offset.

The averaging time was τ = mT, where T was 30 s and m was the number of consecutive spectra in each block. Only points with enough blocks were retained for statistical reliability.

The model-assisted curve approached the shot-noise reference more closely than Lorentzian fitting. The regression did not reduce photon shot noise itself. Its closer approach came from suppressing distortion-induced estimation bias.

Direction, material choice, and current limits

Section titled “Direction, material choice, and current limits”

Directional response remained strong in the fiber-coupled ensemble. A model trained on X-axis spectra predicted X-axis data but produced significantly higher error on Y-axis data. The result shows that a single-orientation response cannot represent transverse components.

Representative spectra from 0.4 to 0.7 mT showed progressive broadening, asymmetric distortion, and shoulder features. They did not separate into well-resolved multiple peaks within the investigated range. The ensemble response remained a broadened spectral distribution rather than a set of individually resolved resonances.

Fiber-coupled ensembles provide a compact probe and improved fluorescence collection, but their random orientations make calibration directional. Such probes suit confined or mobile measurements where bulk geometry is awkward. Single-NV sensors suit localized measurements but have weak fluorescence and strict alignment requirements. Bulk diamond sensors provide stronger stable signals, although their rigid geometry limits confined environments. Ensemble selection therefore depends on spatial scale and deployment constraints, not only NV resonance strength.

The reported method does not establish a general three-axis estimator or performance outside the tested perturbation range. Fisher information served as a relative optimizer, while measured reconstruction error incorporated the actual acquisition noise. Its optimum of approximately 640 µT is tied to the modeled ensemble and fiber geometry.

The ensemble-average model does not describe an isolated NV center. Data underlying the results were not publicly available at publication, although the authors stated that they may be provided on reasonable request.

View Original Abstract

Fiber-coupled ensembles of nanodiamonds containing nitrogen-vacancy (NV) centers are widely used for microscopic magnetic field measurements. However, the random distribution of NV crystal axes in nanodiamond ensembles causes resonance frequencies from differently oriented NV centers to overlap in the ODMR spectrum, distorting the spectral line shape and breaking the linear mapping between frequency shift and magnetic field. As a result, conventional magnetic field measurements based on Lorentzian fitting suffer from significant systematic errors. Here, we establish a fiber-nanodiamond ensemble model to analyze the impact of random NV orientations on the ODMR linear response and define an effective ensemble-averaged linear response coefficient K . Fisher information analysis identifies an information-optimal intrinsic magnetic field of approximately 640 µT, at which the ensemble response can be reliably approximated as linear in its vicinity. Building on this condition, we propose a model-assisted regression scheme, in which the physical fiber-nanodiamond ensemble model defines the optimal operating condition and the associated linear frequency-field scaling, while intrinsic-field-guided Gaussian regression (IFGGR) extracts magnetic-field-induced frequency shifts from distorted ensemble ODMR spectra, enabling accurate magnetic field estimation via the response coefficient K . Under identical acquisition conditions, the proposed method reduces the RMSE from 37.21 µT to 8.99 µT and significantly suppresses systematic bias, while achieving an absolute sensitivity approaching the shot-noise limit at optimal averaging time. The relative slope error reduction (≈15.6-fold) further confirms improved local linearity near the optimal operating point.