## Introduction

Nanoscale magnetic stray fields appearing at surfaces and interfaces of magnetically ordered materials provide important insight into the local spin structure. Such stray fields are present, for example, above magnetic domains and domain walls1, magnetic vortices2, spin spirals3, skyrmions4, or topographic steps and defects5, and often accompany other types of ordering, like ferroelectricity6. Similar stray fields also appear near flowing currents7,8 or materials with a magnetic susceptibility9. Therefore, stray field measurements are a general and versatile tool to study local material or device properties.

While the magnetic imaging of ferromagnetic and ferrimagnetic textures is well-established10,11,12, detection of the much weaker stray fields of, for example, antiferromagnets, multiferroics or nanoscale current distribution is a relatively new development. Quantum magnetometers based on single nitrogen-vacancy (NV) centers have recently led to exciting advances in this direction13,14,15,16,17,18,19. In their standard configuration, NV magnetometers image stray fields by scanning a sharp diamond probe over the sample surface and monitoring the static shift of the NV spin resonance frequency20,21. State-of-the-art scanning NV magnetometers reach a sensitivity to static fields of a few $$\,\mu {{{{{{{\rm{T}}}}}}}}/\sqrt{{{{{{{{\rm{Hz}}}}}}}}}$$16,22. This sensitivity is sufficient for imaging the domain structure of monolayer ferromagnets22,23,24 and uncompensated antiferromagnets14,15,16,17, however, it remains challenging to detect the even weaker stray fields of isolated magnetic defects, spin chains, or ideally compensated antiferromagnets. While higher sensitivities, on the order of $$50\,{{{{{{{\rm{nT}}}}}}}}/\sqrt{{{{{{{{\rm{Hz}}}}}}}}}$$, have been demonstrated using dynamic (ac) detection of fields25,26, this approach is limited to the few systems whose magnetization can be modulated, like isolated spins27,28.

Here, we demonstrate a gradiometry technique for highly sensitive imaging of static magnetization patterns. Our method relies on up-conversion of the local spatial gradient into a time-varying magnetic field using mechanical oscillation of the sensor combined with sensitive ac quantum detection. This operating principle is well known from dynamic force microscopy29,30 and has recently been combined with scanning superconducting quantum interference devices31,32,33. While mechanical oscillation has been explored with NV centers in various forms21,34,35,36,37,38, it has not been realized for sensitive imaging of general two-dimensional magnetic samples. As a striking demonstration, we show that scanning gradiometry is able to resolve the nanotesla magnetic stray fields appearing above single atomic terraces in antiferromagnetic Cr2O3. Imaging of nanoscale current patterns, magnetic susceptibility in metals, and reconstruction of field maps from gradiometry data is also demonstrated.

## Results

The operating principle of our gradiometry technique is shown in Fig. 1a. Our scanning magnetometer set-up consists of a sample plate that is scanned underneath a diamond probe containing a single NV center at the tip apex formed by ion implantation39. The diamond tip is attached to the prong of a quartz tuning fork (Fig. 1) providing atomic force microscopy (AFM) position feedback40. The microscope apparatus additionally includes an objective to optically polarize and readout the NV center spin state and a bond wire acting as a microwave antenna for spin manipulation (see Methods for details). In the conventional mode of operation, we record a spin resonance spectrum at each pixel location to build up a map of the sample’s static stray field16.

To implement the gradiometer, we mechanically oscillate the NV in a plane parallel to the sample (shear-mode) by electrically driving the tuning fork at a fixed amplitude xosc ~ 10–70 nm. The NV center now experiences a time-dependent field given by

$$B(x(t))= \, B({x}_{0})+{\left.\frac{\partial B}{\partial x}\right|}_{x = {x}_{0}}{x}_{{{{{{{{\rm{osc}}}}}}}}}\sin (2\pi {f}_{{{{{{{{\rm{TF}}}}}}}}}t)\\ +{\left.\frac{{\partial }^{2}B}{\partial {x}^{2}}\right|}_{x = {x}_{0}}\frac{{x}_{{{{{{{{\rm{osc}}}}}}}}}^{2}}{2}{\sin }^{2}(2\pi {f}_{{{{{{{{\rm{TF}}}}}}}}}t)+\ldots$$
(1)

where B is the vector component of the sample’s stray field along the NV anisotropy axis (Methods), and where $$x(t)={x}_{0}+{x}_{{{{{{{{\rm{osc}}}}}}}}}\sin (2\pi {f}_{{{{{{{{\rm{TF}}}}}}}}}t)$$ describes the mechanical oscillation around the center location x0 with frequency fTF. The amplitudes of the 0fTF, 1fTF, and 2fTF harmonics in leading orders of xosc are given by

$${B}_{0}=B({x}_{0})$$
(2a)
$${B}_{1}={x}_{{{{{{{{\rm{osc}}}}}}}}}{\left.\frac{\partial B}{\partial x}\right|}_{x = {x}_{0}}$$
(2b)
$${B}_{2}=\frac{1}{2}{x}_{{{{{{{{\rm{osc}}}}}}}}}^{2}{\left.\frac{{\partial }^{2}B}{\partial {x}^{2}}\right|}_{x = {x}_{0}}$$
(2c)

and are therefore proportional to the static field, gradient, and second derivative, respectively. The series expansion in Eq. (2) is accurate for oscillation amplitudes xosc smaller than the scan height, typically d ~ 100 nm. Since the scan height sets the achievable spatial resolution7, the oscillation does not impair imaging resolution.

To detect the harmonics of B(t), we synchronize the mechanical oscillation with a suitable ac quantum sensing sequence41,42, shown in Fig. 2a. Quantum sensing sequences measure the quantum phase accumulated by the coherent precession of a superposition of spin states during an interaction time τ43. To measure the nfTF harmonic, we invert the spin precession n times during one mechanical oscillation period using microwave π-pulses (Carr-Purcell-Meiboom-Gill (CPMG-n) sequence43,44). The pulse protocols for the first (CPMG-1) and second (CPMG-2) signal harmonics are shown in Fig. 2a. The quantum phase accumulated for the first harmonic is given by

$$\phi = \int\nolimits_{t = {t}_{0}-\tau /2}^{{t}_{0}}{\gamma }_{{{{{{{{\rm{e}}}}}}}}}B(t){{{{{{{\rm{d}}}}}}}}t-\int\nolimits_{t = {t}_{0}}^{{t}_{0}+\tau /2}{\gamma }_{{{{{{{{\rm{e}}}}}}}}}B(t){{{{{{{\rm{d}}}}}}}}t\\ = -{\gamma }_{{{{{{{{\rm{e}}}}}}}}}{B}_{1}\,\frac{2{\sin }^{2}(\pi {f}_{{{{{{{{\rm{TF}}}}}}}}}\tau /2)}{\pi {f}_{{{{{{{{\rm{TF}}}}}}}}}}$$
(3)

where γe = 2π × 28 GHz/T is the gyromagnetic ratio of the NV electronic spin and t0 = T/2 (see Fig. 2a). The derivation of Eq. (3) and general expressions for higher harmonics and for sequences that are off-centered with respect to the tuning fork oscillation (t0 ≠ T/2) are given in Supplementary Note 1. To determine ϕ experimentally, we measure the photo-luminescence (PL) intensity CΦ as a function of the phase Φ = x, y, −x, −y of the last microwave π/2 pulse (Fig. 2a). From the four PL signals CΦ we then extract the phase using the two-argument arc tangent (see Methods),

$${\phi }^{({{\mbox{measured}}})}={{\mbox{arctan}}}\,\left(\frac{{C}_{-y}-{C}_{y}}{{C}_{x}-{C}_{-x}}\right)\,.$$
(4)

This four-phase readout technique has the advantage that the phase can be retrieved over the full 2π range with uniform sensitivity19,26. From ϕ(measured) and Eq. (3) we then compute the gradient field B1. Since we can only measure ϕ modulo 2π, a phase unwrapping step is necessary for large signals that exceed the range [−π; π)26. Additionally, the oscillation amplitude can be used as a control parameter (xosc < d) to dynamically set the maximum gradient before phase wrapping occurs.

For NV centers with long coherence times (T2T), the sensitivity can be improved further by accumulating phase over multiple oscillation periods using CPMG-2n sequences (Fig. 2b). The simplest way to meet the T2T condition is to use a mechanical oscillator with a higher resonance frequency. Alternatively, as demonstrated in this work, a higher mechanical mode of the tuning fork can be employed45. High-frequency detection has the added advantage that the interval between π-pulses (given by 1/(2fTF)) becomes very short, which makes dynamical decoupling more efficient and in turn leads to improved sensitivity42,46. We demonstrate single and multi-period detection schemes with sensitivities of $$\sim \!\!120\,{{{{{{{\rm{nT}}}}}}}}/\sqrt{{{{{{{{\rm{Hz}}}}}}}}}$$ and $$\sim\!\!100\,{{{{{{{\rm{nT}}}}}}}}/\sqrt{{{{{{{{\rm{Hz}}}}}}}}}$$, respectively, that exceed our best dc sensitivities of $$\sim\!\!1-2\,\mu {{{{{{{\rm{T}}}}}}}}/\sqrt{{{{{{{{\rm{Hz}}}}}}}}}$$ by one order of magnitude (Supplementary Notes 2 and 3).

We begin measurements by calibrating the oscillation amplitude xosc and verifying Eqs. (2, 3). We characterize the gradiometry phase detection by parking the tip over a fixed position on a magnetic test sample and measuring the B1 and B2 signals as a function of the oscillation amplitude xosc (see Methods and Supplementary Note 4 for characterization and calibration details). Figure 3a confirms that the signals grow as B1xosc and $${B}_{2}\propto {x}_{{{{{{{{\rm{osc}}}}}}}}}^{2}$$, as expected from the Taylor expansion (Eq. 2). Further, to avoid mixing of mechanical and field harmonic signals through surface interactions, we retract the tip by ~20 nm from the contact point while measuring (Supplementary Note 5).

We establish two-dimensional imaging by detecting the stray field from a direct current flowing in a bilayer graphene device (Fig. 3b). This device provides an ideal test geometry because the magnetic stray field and gradient can be tuned and directly compared to the analytical model. Moreover, results can be benchmarked against those obtained by ac current driving26. Figure 3c presents gradiometry maps B1 and B2 and, for comparison, a static field map B0 recorded using a standard dc technique47 with no tuning fork oscillation. Note that all images are recorded with the same imaging time and pixel size. The figure immediately highlights several advantages of gradiometry versus static field imaging: First, the signal-to-noise ratio is strongly enhanced due to the more sensitive ac detection, despite a lower absolute signal. Second, magnetic drift throughout the imaging time (several hours) is present in the static image but suppressed in the gradiometry images. Magnetic drifts are suppressed for two reasons; they fluctuate at low frequencies far away from the detection frequency, and magnetic drift sources do not generate appreciable magnetic field gradients over the oscillation range (~100 nm). Numerically differentiating the static image to remove field drifts is not suitable as it introduces high-frequency noise (Supplementary Fig. 6c). Third, because gradient fields decay quickly with distance (B1x−2 and B2x−3 compared to B0x−1, see Fig. 3d and Supplementary Note 6), they offer a higher apparent image resolution and are thus easier to interpret.

### Imaging of antiferromagnetic surface texture

We now turn our attention to the imaging of stray fields above antiferromagnetic materials, focusing on the archetypical model system Cr2O3. Antiferromagnets represent a general class of weakly magnetic materials that are both challenging to the image by existing techniques48,49,50 and of key importance for understanding multiferroicity and topological magnetism in the context of antiferromagnetic spintronics51. Although antiferromagnets are nominally non-magnetic, weak stray fields can appear due to uncompensated moments at domain walls14,15,16, spin spirals13, topographic steps17, or surface roughness and defects. The latter is particularly difficult to detect, yet plays an important role in the pinning of domain walls17 and establishing an exchange bias at material interfaces52. Careful study of these weaker fields is therefore an important avenue for quantifying the interplay between surface roughness, stray field structure, local magnetization, and domain wall behavior.

In Fig. 4a we show a gradiometry image of a polished Cr2O3 single-crystal16. Cr2O3 has a layered structure of out-of-plane Cr3+ moments at the (0001) surface53,54 that lead to stray fields at topographic steps16,17. These stray fields are proportional to the step height. Indeed, Fig. 4a reveals a rich variety of magnetic anomalies on this surface, including two ~5 nm deep and ~200 nm wide topographic trenches introduced by polishing, a number of point defects, and general texture of the ~2 nm-rms surface roughness (see Supplementary Fig. 11 for topographic characterization). A quantitative calculation of the expected stray field maxima shows that the magnetic anomalies are well explained by the surface topography (Supplementary Note 7). Except for the trenches, these surface defects are not visible in the static field image (Fig. 4b). Note that it is possible to convert the gradiometry map into an improved static field map through integration and weighted averaging in k-space (Fig. 4c and Methods). See Supplementary Fig. 6 for the reverse process, where a B0 image is converted into a B1 image.

Next, we show that gradiometry is sufficiently sensitive to detect single atomic surface steps of the layered Cr2O3(0001) surface. Figure 4d displays a gradiometry image from a second Cr2O3 single crystal with an as-grown (unpolished) surface53. We observe a striking pattern of regular stripes separated by a few hundred nanometers and an approximate amplitude of B1 ~ 250 nT. The pattern extends over the entire crystal surface with different stripe separations and directions in different regions of the sample (Supplementary Fig. 12). By converting B1 measurements into local gradients we deduce height changes <1 nm (Supplementary Note 7), suggesting that the repeating striped patterns are caused by single (h = 0.227 nm), or multiple atomic step edges. To corroborate, we correlate the magnetic features to the AFM topography of the sample (Supplementary Fig. 12). Figure 4e, f show line cuts perpendicular to the stripe direction in two regions of the sample surface, and reveal the presence of both mono- and diatomic step edges. The gradiometry data (Fig. 4g, h), are well fitted by a simple model (see Methods), producing a fitted surface magnetization of σz = 2.1 ± 0.5 μB/nm2, in good agreement with earlier data14,16. Together, our findings unambiguously confirm the stepped growth of the Cr2O3(0001) surface54. To our knowledge, our work reports the first magnetic stray field imaging of atomic steps on an antiferromagnet. It opens a complementary path to existing atomic-scale techniques, such as spin-polarized scanning tunneling microscopy and magnetic exchange force microscopy55, without sharing the need for an ultra-high vacuum, a conducting surface, or cryogenic operation.

### Imaging of magnetic susceptibility

We conclude our study by demonstrating nanoscale imaging of magnetic susceptibility. Susceptometry measurements are important for investigating, for example, the magnetic response of patterned metals and materials9,56, superconductors57 as well as para- and superparamagnetic nanoparticles58,59. Figure 5a shows a gradiometry map of a 50-nm-thick disc made from paramagnetic Pd placed in a bias field of Bext = 35 mT. Under the bias field, the disc develops a magnetization of M = χBext0, where χPd is the magnetic susceptibility of the Pd film. A fit to the data (see Methods), shown in Fig. 5b, produces a susceptibility of χPd = (6.6 ± 0.2) × 10−4. This is slightly smaller than the value of pure Pd (χPd = 7.66 × 10−460). The decreased susceptibility may be attributed to either a finite-size effect61 or hydrogen adsorption62. Repeating the same experiment with Bi, a diamagnetic sample, produces an experimental value of χBi = −(1.7 ± 0.1) × 10−4(Fig. 5c, d), which matches the accepted room temperature value of χBi = − 1.67 × 10−460. Despite the ~4 × weaker susceptibility, the magnetic pattern is clearly visible and its sign is inverted compared to the paramagnetic Pd disc. Additionally, the apparent local structure at the center of the disc is explained by a variation in the film thickness (Supplementary Fig. 13). Together, Fig. 4a–d demonstrate the feasibility of extending sensitive dc susceptometry to the nanometer scale.

## Discussion

Our work demonstrates a simple yet powerful method for imaging static magnetization patterns with high sensitivity and spatial resolution. While we demonstrate scanning gradiometry on a layered antiferromagnet, extension to more challenging systems including perfectly compensated collinear antiferromagnets, screw and step dislocations5, or isolated magnetic defects is natural. In particular, the magnetic signal from an atomic step edge is equivalent to that of a one-dimensional spin chain with a linear magnetization density of σzh ≈ 0.5 μB/nm (Methods), demonstrating the feasibility of imaging generic 1D spin systems. Gradiometry is also well-positioned for imaging the internal structure of domain walls and skyrmions, and in particular, for quantifying their size and chirality16,63,64,65 (Supplementary Fig. 14). Finally, while we demonstrate gradiometry on magnetic fields, the technique can be extended to electric fields by orienting the external bias magnetic field perpendicular to the NV axis66,67,68. In particular, the dynamic mode of operation may alleviate charge screening that has previously limited scanning NV electrometry of dc electric field sources69. With the ability to image both magnetic and electric fields, one could imagine correlating antiferromagnetic and ferroelectric order in multiferroics51, providing a unique angle to investigate the magneto-electric coupling in these fascinating and technologically important materials.

## Methods

### Experimental set-up

All experiments were carried out at room temperature with two home-built scanning microscopes. Micro-positioning was carried out by closed-loop three-axis piezo stages (Physik Instrumente) and AFM feedback control was carried out by a lock-in amplifier (HF2LI, Zurich Instruments) and standard PID controls. PL of the NV centers was measured with avalanche photodiodes (APDs) (Excelitas) and data were collected by standard data acquisition cards (PCIe-6353, National Instruments). Direct currents sent through the bilayer graphene were created with an arbitrary waveform generator (DN2.663-04, Spectrum Instrumentation) and the current was measured with a trans-impedance amplifier (HF2TA, Zurich Instruments). Microwave pulses and sequences were created with a signal generator (Quicksyn FSW-0020, National Instruments) and modulated with an IQ mixer (Marki) and an arbitrary waveform generator (DN2.663-04, Spectrum Instrumentation, and HDAWG, Zurich Instruments). NV centers were illuminated at < 100 μW by a custom-designed 520 nm pulsed diode laser. Scanning NV tips were purchased from QZabre AG39. Three different NV tips were used throughout this study with standoff distances d between 70 and 130 nm (excluding the 20-nm retract distance). When imaging our samples, we typically measured for 5 to 30 seconds per pixel to achieve good signal-to-noise. We observed relatively small spatial drifts (typically <30-nm per day). As a result, we did not employ any drift correction techniques and we do not expect drifts to affect our results or scientific claims.

### Magnetic samples

#### Bilayer graphene device

Standard microfabrication processes were used, including mechanical exfoliation and a dry transfer process for generating the hBN-bilayer-graphene-hBN stack, electron beam lithography, and plasma etching for creating the device geometry and physical vapor deposition for creating the metallic device contacts. Full details can be found in ref. 26.

#### Cr2O3 crystals

The mechanically polished and non-polished Cr2O3 bulk single crystals were provided by professor Manfred Fiebig. Both crystals have a (0001) surface orientation. Crystal growth and processing details can be found in refs. 16,53.

#### Pd and Bi discs

Metallic micro-discs were created using standard microfabrication processes. Discs were defined through electron beam lithography on spin-coated Si wafers. Chemical development of the spin-coated resist, followed by metal deposition of either 50-nm of Pd or Bi, and a lift-off process finished the fabrication.

### Initialization and readout of NV spin state

A laser pulse of ~2 μs duration was used to polarize the NV center into the mS = 0 state. Then, the dc or ac quantum sensing measurement occurred on the mS = 0 to mS = −1 transition (see below). The readout of the NV’s spin state was performed by another ~2 μs long laser pulse, during which the photons emitted from the NV center were collected with the APD, binned as a function of time, and summed over a window (~300 ns) that optimized spin state-dependent PL contrast70.

### DC sensing protocol

DC magnetic images were acquired with the pulsed ODMR method47. Measured magnetic resonance spectra were fitted for the center frequency fc of a Lorentzian function, $$L(f)=1-\epsilon {[(2\pi f-2\pi {f}_{c})/{\omega }_{L}]}^{2}+1{\left.\right]}^{-1}$$, where ϵ is the spin contrast (in percent) and ωL is the width of the Lorentzian dip. In a 2D scan the magnetic field projected along the axis of the NV could be determined at each pixel using B(x, y) = 2π[fc(x, y) − f0]/γe where f0 is the resonance frequency far from the surface. For our diamond probes, the NV anisotropy axis was at an θ ~ 55 angle with respect to the out-of-plane direction (z-axis in Fig. 1). Therefore, B(x, y) corresponded to the vector field projection along this tilted direction.

### AC sensing protocol

AC sensing used either a spin echo or dynamical decoupling sequence41,42. The NV spin state was initialized optically into the mS = 0 state followed by a microwave π/2-pulse to create a coherent superposition between the mS = 0 and mS = −1 states. The quantum phase ϕ accumulated between the two states during the coherent precession can be expressed as $$\phi =\int\nolimits_{0}^{\tau }{\gamma }_{{{{{{{{\rm{e}}}}}}}}}g(t)B(t)dt$$ where g(t) is the modulation function43, B(t) is the magnetic field, τ is the interaction time (the time in between the first and last π/2-pulse), and where we use the rotating frame approximation. The modulation function alternates between ± 1 with each microwave π-pulse during the pulse sequence. While imaging with the gradiometry pulse sequences, 3π/2-pulses were sometimes used instead of the final π/2-pulses for the projective spin readout to reduce pulse imperfections caused by (ca. ± 100 kHz) drifts in the NV resonance frequency.

### Calibration of tuning fork oscillation

We estimated the oscillation amplitude xosc and oscillation angle in the xy-plane, denoted by α, of the tuning forks with two different in-situ measurement techniques. The first method involved processing a static field map and a gradient field map acquired over the same region with a least-squares minimization scheme. By minimizing a cost function proportional to the pixel value differences between the B1 and numerically differentiated B0 images, estimates for xosc and α were determined. The second method involved stroboscopic imaging of the static field (B0) synchronized to the tuning fork oscillation. By measuring the time-tagged displacements of magnetic features recorded at different positions during the tuning fork oscillation, xosc and α could be estimated by fitting to the path taken by the NV. We found that xosc scaled linearly with the applied drive voltage for the voltages we used while imaging. Examples of these calibration methods are given in Supplementary Note 4.

### Calibration of the trigger delay

In order to optimize the ac sensing sequences, and to distinguish between signals from the first and second derivatives (see Supplementary Note 1), a trigger delay (t0) calibration measurement must be made. The calibration measurement consisted of measuring the phase ϕ at a stationary point on the sample, while varying t0, thus producing a phase that oscillated as a function of t0. The chosen value of t0 was the value that maximized the measured phase. Examples of this calibration measurement, as well as the characterization of PL oscillation caused by the tuning fork motion, are shown in Supplementary Note 4.

### Reconstruction of the static field map from the gradient map

A gradient map B1 can be used to reduce the noise and improve the image of a (less sensitive) static field map B0. While direct integration of B1 produces a B0 map32, a combination approach using B0 and B1 provides the best noise suppression across all spatial wavelengths, with no need for ad-hoc boundary conditions. Letting BT denote the true static magnetic field value, the experimentally measured B0 and B1 field maps are

$${B}_{0}(x,y) ={B}_{{{{{{{{\rm{T}}}}}}}}}(x,y)+{w}_{0}\\ {B}_{1}(x,y) ={x}_{{{{{{{{\rm{osc}}}}}}}}}\frac{\partial }{\partial r}{B}_{{{{{{{{\rm{T}}}}}}}}}(x,y)+{w}_{1}$$
(5)

where $$\frac{\partial }{\partial r}=\cos (\alpha )\frac{\partial }{\partial x}+\sin (\alpha )\frac{\partial }{\partial y}$$ is the directional derivative and $${w}_{0} \sim {{{{{{{\mathcal{N}}}}}}}}(0,{\sigma }_{{B}_{0}}^{2})$$ and $${w}_{1} \sim {{{{{{{\mathcal{N}}}}}}}}(0,{\sigma }_{{B}_{1}}^{2})$$ are white noise added to each pixel (reflecting the Poissonian shot noise of the photo-detection). In k-space these equations can be transformed into

$${\hat{B}}_{0}({k}_{x},{k}_{y}) ={\hat{B}}_{{{{{{{{\rm{T}}}}}}}}}({k}_{x},{k}_{y})+{\hat{w}}_{0}\\ \frac{{\hat{B}}_{1}({k}_{x},{k}_{y})}{i{x}_{{{{{{{{\rm{osc}}}}}}}}}{k}_{r}} ={\hat{B}}_{{{{{{{{\rm{T}}}}}}}}}({k}_{x},{k}_{y})+\frac{{\hat{w}}_{1}}{i{x}_{{{{{{{{\rm{osc}}}}}}}}}{k}_{r}}$$
(6)

where $$\hat{X}$$ denotes the Fourier transform of X, $${k}_{r}={k}_{x}\cos \alpha +{k}_{y}\sin \alpha$$ is the dot product between the oscillation direction and the k-vector [kx, ky]. Integration in k-space introduces a k-dependent noise term in the gradient field map. In particular, the integrated gradient map has noise amplification near the line $${k}_{y}=-\cot (\alpha ){k}_{x}$$ but has noise suppression far away from that line. Directional sensitivity could be avoided by oscillating the tuning fork in the z-direction (tapping mode), or by using an oscillator that supports orthogonal lateral modes. To circumvent this problem we average two k-space maps (Eq. (6)) with k-vector dependent weights that reflect the noise added by the integration process. The use of inverse variance weights additionally results in an image with the lowest possible variance. Thus, the optimal reconstructed static field map can be computed by taking the inverse Fourier transform of

$${\hat{\tilde{B}}}_{0}({k}_{x},{k}_{y})={\hat{B}}_{0}({k}_{x},{k}_{y})\,\frac{{k}_{0}^{2}}{{k}_{0}^{2}+{k}_{r}^{2}}+\frac{\hat{{B}}_{1}({k}_{x},{k}_{y})}{i{x}_{{{{{{{{\rm{osc}}}}}}}}}{k}_{r}}\,\frac{{k}_{r}^{2}}{{k}_{0}^{2}+{k}_{r}^{2}}$$
(7)

where $${k}_{0}^{2}={\sigma }_{{B}_{1}}^{2}/({x}_{{{{{{{{\rm{osc}}}}}}}}}^{2}{\sigma }_{{B}_{0}}^{2})$$ defines a cut-off wave vector determined by the oscillation amplitude and the noise variances $${\sigma }_{{B}_{0}}^{2}$$ and $${\sigma }_{{B}_{1}}^{2}$$. The corresponding cut-off wavelength λ = 2π/k0 reflects the spatial wavelength above (below) which the B0 (B1) field map is less noisy. It should also be noted that the k-space averaging process can be modified to include the B2 map, however, the B1 map provides the most significant improvement. Specifically, in Fig. 4c, xosc = 23 nm and α = 180 were used as reconstruction parameters (determined by the calibration in Supplementary Note 4). Note, since the gradient and k-space averaging are directional, the noise reduction is also directional (approximately 22 × in the x-direction and 5.6 × in the y-direction).

### Stray fields from atomic step edges

For an atomic step edge propagating along the y-direction the stray field is modeled by two out-of-plane magnetic samples with different heights. Taking the analytical form (see ref. 63) for the stray fields above an edge at x = 0 as $${B}_{x}(x,z)=\frac{-{\mu }_{0}{\sigma }_{z}}{2\pi }\frac{z}{{x}^{2}+{z}^{2}}$$, By = 0, and $${B}_{z}(x,z)=\frac{{\mu }_{0}{\sigma }_{z}}{2\pi }\frac{x}{{x}^{2}+{z}^{2}}$$, where σz is the surface magnetization, we define the stray field produced by a small change in the height h as

$$\begin{array}{ll}{B}_{x}^{{{{{{{{\rm{step}}}}}}}}}&={B}_{x}(x,d)-{B}_{x}(x,d+h)\approx \frac{{\mu }_{0}{\sigma }_{z}h}{2\pi }\frac{({x}^{2}-{d}^{2})}{{({x}^{2}+{d}^{2})}^{2}}\\ {B}_{z}^{{{{{{{{\rm{step}}}}}}}}}&={B}_{z}(x,d)-{B}_{z}(x,d+h)\approx \frac{{\mu }_{0}{\sigma }_{z}h}{2\pi }\frac{2xd}{{({x}^{2}+{d}^{2})}^{2}}\end{array}$$
(8)

with $${B}_{y}^{{{{{{{{\rm{step}}}}}}}}}=0$$. The expressions are simplified in the limit of hd since the sub-nanometer atomic step edges are much smaller than typical standoff distances of d ~ 50 − 100 nm. We measure the gradient of the stray field along the oscillation direction, taken as the x-direction for simplicity. This leads to field gradients of

$$\frac{\partial {B}_{x}^{{{{{{{{\rm{step}}}}}}}}}}{\partial x}=\frac{{\mu }_{0}{\sigma }_{z}h}{\pi }\;\frac{x(3{d}^{2}-{x}^{2})}{{({x}^{2}+{d}^{2})}^{3}}\\ \frac{\partial {B}_{z}^{{{{{{{{\rm{step}}}}}}}}}}{\partial x}=\frac{{\mu }_{0}{\sigma }_{z}h}{\pi }\;\frac{d({d}^{2}-3{x}^{2})}{{({x}^{2}+{d}^{2})}^{3}}$$
(9)

The fits in Fig. 4g, h are produced by projecting the field gradients onto the NV axis defined by $${{{{{{{\bf{e}}}}}}}}=[\sin \theta \cos \varphi ,\sin \theta \sin \varphi ,\cos \theta ]$$ and summing over multiple steps. The line cuts are produced by rotating and plane averaging a 2D image (Supplementary Fig. 12). We account for the rotation angle in the xy-plane by introducing an additional image rotation angle $$\varphi ^{\prime}$$. The experimentally measured B1 field is then fitted by

$$\frac{{B}_{1}}{{x}_{{{{{{{{\rm{osc}}}}}}}}}\cos (\varphi ^{\prime} )}=\mathop{\sum}\limits_{{{{{{{{\rm{steps}}}}}}}}}\sin (\theta )\cos (\varphi +\varphi ^{\prime} )\frac{\partial {B}_{x}^{{{{{{{{\rm{step}}}}}}}}}}{\partial x}+\cos (\theta )\frac{\partial {B}_{z}^{{{{{{{{\rm{step}}}}}}}}}}{\partial x}$$
(10)

We set h = 0.227 nm and $$\varphi ^{\prime} =4{0}^{\circ }$$ for the fit in Fig. 4g, h = 0.454 nm and $$\varphi ^{\prime} =2{3}^{\circ }$$ for the fit in Fig. 4h and fixed xosc = 46 nm (determined by the tip calibration) in both fits. The standoff distance d, the surface magnetization σz, angles θ and φ and step edge locations are free parameters in the fit. Collectively, the fitted line scans give d = 89 ± 12 nm (including the 20 nm retract distance) and σz = 2.1 ± 0.5 μB/nm2, which is consistent with previous measurements on this sample16.

Note, Eq. (8) is functionally equivalent to the magnetic field from a one-dimensional ferromagnetic spin chain with a linear magnetization density of M1D = σzh. To demonstrate this, we calculate the stray field produced by a infinite line of magnetic dipoles along the line x = 0 and pointing in the x-direction as:

$${B}_{x}^{1{{{{{{{\rm{D}}}}}}}}}= \,\frac{{\mu }_{0}{M}^{1{{{{{{{\rm{D}}}}}}}}}}{4\pi }\int\nolimits_{-\infty }^{\infty }\left(\frac{3{x}^{2}}{{r}^{5}}-\frac{1}{{r}^{3}}\right)dy=\frac{{\mu }_{0}{M}^{1{{{{{{{\rm{D}}}}}}}}}}{2\pi }\frac{({x}^{2}-{d}^{2})}{{({x}^{2}+{d}^{2})}^{2}}\\ {B}_{z}^{1{{{{{{{\rm{D}}}}}}}}}= \,\frac{{\mu }_{0}{M}^{1{{{{{{{\rm{D}}}}}}}}}}{4\pi }\int\nolimits_{-\infty }^{\infty }\frac{3xd}{{r}^{5}}dy=\frac{{\mu }_{0}{M}^{1{{{{{{{\rm{D}}}}}}}}}}{2\pi }\frac{2xd}{{({x}^{2}+{d}^{2})}^{2}}$$
(11)

where r2 = x2 + y2 + d2.

### Susceptibility fitting for Pd and Bi discs

Susceptibility fits followed the assumption that the stray field produced by a para- or diamagnetic sample is identical to that of a homogeneously magnetized body with a magnetization magnitude of M = χPd/BiBext0 and a magnetization vector that is parallel to the external polarizing field. The stray field produced by this magnetization is B(r) = − μ0ϕmag(r) where ϕmag(r) is the magnetic potential. We computed the stray field using the k-space method of ref. 71 and fitted the gradient image to the numerically computed gradient fields. In the fitting procedure we fixed xosc = 69 nm, α = 180 (determined by the tip calibration) and sample thickness t = 50 nm while the standoff distance d, magnetization magnitude M, angles θ and φ, circle radius and position are free parameters in the fit. The susceptibilities are computed as χPd/Bi = μ0M/Bext and the fitted magnetizations were MPd = 18.6 ± 0.6 A/m and MBi = −4.4 ± 0.4 A/m.