betaFeedback

WorksPhysics and Astronomy

Physics-informed X-ray tomography of magnetization and material parameters

Abstract

Three-dimensional spin textures are stabilized by microscopic interactions within their host materials. X-ray vector tomography can image these nanoscale structures, but weak magnetic contrast and a limited range of viewing angles often leave parts of the magnetization poorly determined. Here we show that a reconstruction constrained by the micromagnetic energy recovers the three-dimensional magnetization from such data and gains sensitivity to the underlying material interactions. Applied to the chiral magnet Co8_8Zn9_9Mn3_3, the method resolves the three-dimensional helical texture even from a small fraction of the dual-axis projections typically required for a successful tomographic reconstruction, or from a single tilt axis alone. By testing which ratio of exchange to Dzyaloshinskii-Moriya interaction in the reconstruction best reproduces the measured projections, we obtain a value for the sample that differs by about 10 % from the literature value derived from the helical winding period. Taken together, our findings suggest a way to reduce the time needed for vector tomography and to extend it to determining fundamental magnetic interactions in a material.

Magnetic materials support a rich family of three-dimensional textures, including skyrmion strings [1, 2], hopfions [3], and vortex rings [4]. These textures carry a nontrivial topology that governs how they form, move, and interact, which makes them candidates for three-dimensional spintronic devices [5, 6]. To understand the three-dimensional structure and behavior of these textures, we must image the full magnetization vector field. X-ray magnetic vector tomography provides this capability by reconstructing the vector field from projections recorded with X-ray magnetic circular dichroism (XMCD) contrast [7] about two rotation axes [8, 9, 10]. X-ray vector imaging has since been extended to laminography [11], and electron magnetic tomography has been used to reconstruct magnetic vector fields from holographic [12] or Lorentz [13] images. However, textures that vary across the depth as well as in the plane require samples several hundred nanometers or even micrometers thick. To study these systems, high-resolution imaging has been made possible with the use of hard X-rays [8] and pre-edge phase contrast with coherent soft X-ray techniques [14]. However, weak magnetic scattering can result in noisy dichroic projections. This, combined with practical constraints on accessible measurement angles, radiation dose, and available synchrotron beamtime, can result in nanoscale features in the magnetization vector field being poorly constrained [15].

To prevent noise from distorting the reconstruction, we can spatially smooth it by penalizing sharp changes in the magnetization [16, 17]. Mathematically, this gradient penalty has the same form as the exchange interaction (stiffness AA) that aligns neighboring moments in micromagnetic theory [18, 19], giving the smoothing a direct physical basis. However, because the experimental resolution often approaches the scale of the magnetic features, this smoothing alone blurs the domain walls and helical winding we may want to image. In micromagnetic theory, stable states are local minima of the full energy functional, of which exchange is only one term. In a real magnet, these features persist because exchange competes with the demagnetizing interaction (scaled by the saturation magnetization MSM_{\mathrm{S}}), which favors flux closure, and in chiral magnets also with the Dzyaloshinskii–Moriya interaction (constant DD), which twists neighboring moments. Previous studies have applied the micromagnetic model only after the reconstruction, adjusting simulations by hand to reproduce the imaged texture [2, 12, 20]. This manual tuning is time-consuming and cannot restore features that the reconstruction has already lost or distorted.

Here, we present an approach that avoids this loss of information by constraining the reconstruction itself with the micromagnetic energy, similar to our approach when recovering the magnetization of a thin film from stray-field magnetometry [21]. We extend X-ray magnetic vector tomography with the complete micromagnetic energy functional of the sample and demonstrate its two roles on a thick disk of the room-temperature chiral magnet Co8Zn9Mn3\mathrm{Co}_8\mathrm{Zn}_9\mathrm{Mn}_3 [22], imaged with dichroic soft X-ray ptychography. First, the energy encodes our prior knowledge of the interactions in the material and helps us recover the magnetization vector field where a limited number of noisy projections alone is not enough. Second, because the reconstruction explicitly incorporates these competing interactions, it allows us to test the assumed model itself. Since material parameters are usually only determined for the bulk and do not always translate to a nanostructure, we identify the relevant parameters of the sample that best reproduce the measured projections. This prospect of texture-driven magnetometry on individual nanostructures is relevant for a wide variety of magnetic material systems.

Physics-informed reconstruction from XMCD projections

In magnetic tomography, XMCD imaging records two projections at every sample orientation, with opposite circular X-ray polarization σ±\sigma^{\pm}, giving a transmission image pair C+C^{+} and C−C^{-} [8, 23]. At the X-ray absorption edges of nickel, the XMCD effect provides contrast that is sensitive to the local magnetization component along the beam propagation direction [7] and reverses sign between σ+\sigma^{+} and σ−\sigma^{-}. We sum C+C^{+} and C−C^{-} to yield a purely electronic channel cc and subtract them to yield a magnetic (XMCD) channel yy (Figure 1b and Methods). Because XMCD contrast is proportional to the magnetization component parallel to the beam propagation direction, each measured projection yiy_i records the line integral of that component along the beam [8]. In our geometry, the beam stays along z^\hat{z} while the stage rotates the sample by Ri\mathbf{R}_i (Figure 1a), so in the sample frame the beam direction is k^i=Riz^\hat{\mathbf{k}}_i=\mathbf{R}_i\hat{\mathbf{z}} [9]. The forward model simulates the measurement of a given magnetization configuration: for normalized magnetization m=M/Ms\mathbf{m} = \mathbf{M}/M_{\mathrm{s}}, it computes the simulated projection pi(m)\mathbf{p}_i(\mathbf{m}), the counterpart of the measured yi\mathbf{y}_i, via the discrete Radon transform Ri\mathcal{R}_i, the operator that takes these line integrals [24]:

Physics-informed magnetization reconstruction from XMCD tomography

Figure 1. Physics-informed magnetization reconstruction from XMCD tomography. a, Dual-axis tomography. Circularly polarized X-rays transmit through the sample, which we rotate about y^\hat{y} by θxz∈[−41∘,+15∘]\theta_{xz} \in[-41^\circ,+15^\circ] (xzxz series) and about x^\hat{x} by θyz∈[−30∘,+47∘]\theta_{yz} \in[-30^\circ,+47^\circ] (yzyz series). b, X-ray imaging. At every tilt we record transmission images at both circular polarizations σ±\sigma^{\pm}. The pair C±C^{\pm} separates into the magnetic XMCD channel yy and the non-magnetic electronic channel cc [Eq. (5)]. c, The measured projections yiy_i feed into the physics-informed reconstruction, which minimizes the joint objective L=Ldata2+λEEtotal\mathcal{L}=\mathcal{L}_{\mathrm{data}}^2+\lambda^E E_{\mathrm{total}} [Eqs. (2) and (4)], shown at λE=1014\lambda^E=10^{14}. Gradient-based optimization returns the reconstruction m⋆\mathbf{m}^{\star}, shown by its components mim_i (i=x,y,zi=x,y,z). Scale bars, 200 nm.

pi(m)∝Ri[m⋅k^i].(1)\mathbf{p}_i(\mathbf{m}) \propto\mathcal{R}_i\left[\mathbf{m}\cdot\hat{\mathbf{k}}_i\right]. \tag*{(1)}

Unlike the scalar density imaged in conventional tomography, the magnetization is a vector field probed only along the beam, so rotating the sample changes not only the viewing angle but also which component is sensed. To reconstruct all three vector components, the acquisition therefore needs to combine two orthogonal tilt series (Figure 1a) or to use a laminography geometry [11]. Tilting about y^\hat{\mathbf{y}} confines the beam to the xzxz plane, so this series (the xzxz series) is sensitive to mxm_x and mzm_z but never to mym_y. Tilting about x^\hat{\mathbf{x}} confines the beam to the yzyz plane, so the yzyz series is sensitive to mym_y and mzm_z but never to mxm_x.

Recovering the three-dimensional magnetization configuration requires finding a vector field whose forward projections reproduce the measured data (Figure 1c). Even a complete dual-axis acquisition does not uniquely determine the magnetization: line integrals of m⋅k^\mathbf{m}\cdot\hat{\mathbf{k}} determine at most the divergence-free part of the magnetization [25, 26], since any internal gradient field that vanishes at the sample boundary integrates to zero along every beam path [9]. In physical terms, the magnetic volume and surface charges of the sample cannot be fully recovered without additional constraints. In practice, depending on the sample geometry and the setup, projections cannot always be recorded over the complete angular range, leaving a so-called missing wedge of unrecorded viewing directions in tomography and a missing cone in laminography [27]. Moreover, weak XMCD signals make the projections particularly susceptible to noise and imaging artifacts across the angular range.

Despite these challenges, iterative magnetic vector tomography [8, 9] has recovered three-dimensional magnetic textures by minimizing the data misfit between simulated and measured projections. However, without further constraints, these limitations can leave nanoscale features poorly resolved and introduce artifacts into the reconstruction. We therefore add the micromagnetic energy of the material to this data misfit and minimize the combined objective function L(m)\mathcal{L}(\mathbf{m}),

L(m)=Ldata2(m)+λEEtotal(m),(2)\mathcal{L}(\mathbf{m}) = \mathcal{L}_{\mathrm{data}}^{2}(\mathbf{m}) + \lambda^{E} E_{\mathrm{total}}(\mathbf{m}), \tag*{(2)}

where the data misfit Ldata\mathcal{L}_{\mathrm{data}}, a normalized form of the error metric of magnetic vector tomography [9], measures how closely the simulated projections reproduce the NN measured projections yi\mathbf{y}_i,

Ldata2(m)=1N∑i=1N∥pi(m)−yi∥2∥yi∥2.(3)\mathcal{L}_{\mathrm{data}}^{2}(\mathbf{m}) = \frac{1}{N}\sum_{i=1}^{N}\frac{\lVert\mathbf{p}_i(\mathbf{m})-\mathbf{y}_i\rVert^{2}}{\lVert\mathbf{y}_i\rVert^{2}}. \tag*{(3)}

The energy weight λE\lambda^{E} balances this data misfit against the micromagnetic energy.

The micromagnetic energy enters Eq. (2) as the physics prior. This penalty increases with the energy of the magnetization state, driving the reconstruction toward physically plausible configurations. It sums the exchange, demagnetizing, and bulk Dzyaloshinskii–Moriya (DMI) energies,

Etotal(m)=Eex+Edem+Edmi,(4)E_{\mathrm{total}}(\mathbf{m}) = E^{\mathrm{ex}} + E^{\mathrm{dem}} + E^{\mathrm{dmi}}, \tag*{(4)}

parameterized by the exchange stiffness AA, saturation magnetization MsM_{\mathrm{s}}, and DMI constant DD Eq. (8), Methods. The micromagnetic energy is only meaningful inside the material, where the magnetization has unit length (∣m∣=1\lvert\mathbf{m}\rvert= 1), whereas the surrounding vacuum carries no magnetization at all. We therefore define a disk mask ww that marks which cells of the reconstruction volume belong to the sample, and restrict both the forward model and the energy computation to these cells. Previous studies defined such a mask from a scalar tomogram of the non-magnetic electronic channel [8]. For our flat disk, however, the missing wedge elongates this tomogram along zz (Methods). We instead take the disk outline from the electronic channel of the face-on projection and extend it through the fabricated 300 nm thickness. We minimize L\mathcal{L} using gradient-based optimization, with automatic differentiation supplying the gradients [28].

Reconstructing the magnetic state

To demonstrate our physics-informed reconstruction, we apply it to tomographic data measured on a 1.1 μ\mum-diameter, 300 nm-thick Co8_8Zn9_9Mn3_3 disk to recover its three-dimensional magnetization configuration. Co8_8Zn9_9Mn3_3 belongs to the Co–Zn–Mn family of room-temperature helical magnets [22], which host skyrmions [22], Bloch points [20], and hopfions [29]. We image the disk in the remanent state, where it hosts a helical configuration. The XMCD phase projections were measured 3 eV below the Co L3L_3 absorption edge (778.1 eV) using dichroic soft X-ray ptychography, a coherent diffractive imaging method that recovers the phase shift of the transmitted X-rays. At this pre-edge energy, the X-rays are only weakly attenuated while the magnetic phase contrast persists, which allows imaging of samples up to an order of magnitude thicker than at the absorption edge [14]. The two tilt series cover θxz∈[−41∘,+15∘]\theta_{xz}\in[-41^\circ,+15^\circ] (29 projections, xzxz series) and θyz∈[−30∘,+47∘]\theta_{yz}\in[-30^\circ,+47^\circ] (38 projections, yzyz series) and a pixel size of about 10 nm. To reconstruct the magnetization with Eq. (2), we choose the energy weight λE\lambda^{E} and specify the material parameters. We use the literature values of AA and MsM_{\mathrm{s}} [20] and D=0.66×10−3 J m−2D=0.66\times10^{-3}\,\mathrm{J\,m^{-2}}, the DMI constant that best reproduces the measured projections (Sec. IV, Extended Data Table I).

We first examine how the energy weight λE\lambda^{E} influences the reconstructed state by sweeping λE\lambda^{E} across several orders of magnitude and plotting the micromagnetic energy EtotalE_{\mathrm{total}} against the data misfit Ldata\mathcal{L}_{\mathrm{data}} in an L-curve (Figure 2a), similar to how regularization weights are determined in inverse problems [30, 31]. At λlowE\lambda^{E}_{\mathrm{low}}, the physics prior is too weak to constrain the reconstruction, which is nearly identical to an unconstrained one (λE=0\lambda^{E} = 0, Extended Data Figure 1) and shows a limitation common to magnetic vector tomography: because the X-ray beam stays close to z^\hat{z} over the restricted tilt range, the measurement strongly constrains the out-of-plane stripes in mzm_z, whereas the weakly probed in-plane components mxm_x and mym_y remain much noisier and less resolved (Figure 2b, top row). At λhighE\lambda^{E}_{\mathrm{high}}, the energy term pulls the reconstruction toward a low-energy state at the expense of data agreement: the mzm_z stripes differ from both the λlowE\lambda^{E}_{\mathrm{low}} reconstruction and the measured projections (Figure 2b, bottom row). To select the balanced weight λbalE\lambda^{E}_{\mathrm{bal}}, we increase λE\lambda^{E} up to the bend of the L-curve (λbalE=1014\lambda^{E}_{\mathrm{bal}} = 10^{14}), just before the reconstruction begins to clearly deviate from the measurements (Ldata=0.83\mathcal{L}_{\mathrm{data}} = 0.83 versus 0.780.78). At this weight, mzm_z keeps the stripe pattern of the λlowE\lambda^{E}_{\mathrm{low}} reconstruction and of the measured projections (Extended Data Figure 2), and mxm_x and mym_y form a coherent three-dimensional helical texture (Figure 2b, middle row).

Figure 2. Influence of the energy weight λE\lambda^{E} on the reconstruction. a, L-curve of micromagnetic energy EtotalE_{\mathrm{total}} versus data misfit Ldata\mathcal{L}_{\mathrm{data}} as λE\lambda^{E} is swept, marking three reconstructions with increasing weight: a low weight (λlowE=1012\lambda^{E}_{\mathrm{low}} = 10^{12}), the balanced weight (λbalE=1014\lambda^{E}_{\mathrm{bal}} = 10^{14}), and a high readout weight (λhighE=1020\lambda^{E}_{\mathrm{high}} = 10^{20}, Sec. IV). b, Reconstructed magnetization components mxm_x, mym_y, and mzm_z at these three operating points. Scale bar, 200 nm.

Extended Data Figure 1. Synthetic validation of both operating regimes. Synthetic projections are generated from a known magnetization state with noise and reconstructed with the experimental protocol (Methods, ’Synthetic validation’). a, The reconstruction of the measured disk without the energy term (λE=0\lambda^{E}=0, top) and the ground truth relaxed from it (bottom), components mxm_x, mym_y, mzm_z in the disk plane. b, Input XMCD projections, noiseless (top) and with noise (bottom), for three views: the xzxz series at its largest tilt, normal incidence, and the yzyz series at its largest tilt. c, The reconstructions of the four acquisitions of Figure 3 at λE=1015\lambda^{E}=10^{15}, laid out as Figure 3b. Each row label gives the relative error εm\varepsilon_m to the truth. d, L-curve of the full acquisition, micromagnetic energy EtotalE_{\mathrm{total}} versus data misfit Ldata\mathcal{L}_{\mathrm{data}} for λE\lambda^{E} from 101210^{12} to 102010^{20}, with λbalE\lambda^{E}_{\mathrm{bal}} and λhighE\lambda^{E}_{\mathrm{high}} indicated. e, εm\varepsilon_m of the full acquisition versus λE\lambda^{E}, with λbalE\lambda^{E}_{\mathrm{bal}}, λhighE\lambda^{E}_{\mathrm{high}} and the value at λE=0\lambda^{E}=0 indicated. f, Reconstructions with AA, DD, or MsM_s set to a wrong value, in the balanced regime (λbalE\lambda^{E}_{\mathrm{bal}}, left) and the readout regime (λhighE=1020\lambda^{E}_{\mathrm{high}}=10^{20}, right, DD on the refined grid of Figure 4a). The top row shows εm\varepsilon_m and the bottom row the data misfit Ldata\mathcal{L}_{\mathrm{data}}. Scale bars, 200 nm.

Extended Data Figure 2. Measured, modeled and residual projections. Measured XMCD projections yy (top row of each block), the projections p(m⋆)p(\mathbf{m}^{\star}) of the reconstruction at λbal\lambda_{\mathrm{bal}} (middle) and the residual y−p(m⋆)y-p(\mathbf{m}^{\star}) (bottom), for the xzxz series (upper block) and the yzyz series (lower block), on one gray scale. Each block shows five single frames spread evenly over the tilt range in which the stripes are resolved: normal incidence, the ends of that range, and the frames halfway to either end. Numbers in the residual panels give the frame-wise relative residual ∥pi(m⋆)−yi∥/∥yi∥\lVert p_i(\mathbf{m}^{\star})-y_i\rVert/\lVert y_i\rVert, whose root-mean-square over all NN frames is Ldata=0.83\mathcal{L}_{\mathrm{data}}=0.83 (Eq. (3)). Scale bar, 200 nm.

However, the true magnetization of the measured disk is unknown, and the data misfit alone cannot show whether the reconstructed mxm_x and mym_y are correct. We therefore test the method on a synthetic disk whose magnetization mGT\mathbf{m}^{\mathrm{GT}} we know [9]. To obtain a realistic mGT\mathbf{m}^{\mathrm{GT}}, we start from the unconstrained reconstruction (λE=0\lambda^{E} = 0) of the measured disk and relax it into an energy minimum with the Landau–Lifshitz–Gilbert equation [32, 33] in NeuralMag [28] (Extended Data Figure 1). We then simulate the experiment on this disk: we compute its projections at the measured tilt angles, add noise matched to the measurement, and reconstruct the magnetization from these synthetic projections with the same physics-informed protocol. Comparing the reconstruction m∗\mathbf{m}^{*} with the known mGT\mathbf{m}^{\mathrm{GT}} gives the relative error ϵm=∥m∗−mGT∥/∥mGT∥\epsilon_m = \lVert\mathbf{m}^{*} - \mathbf{m}^{\mathrm{GT}}\rVert/\lVert\mathbf{m}^{\mathrm{GT}}\rVert. Without the physics prior (λE=0\lambda^{E} = 0), the reconstruction recovers the mzm_z stripes but fails to resolve the in-plane components, resulting in a large overall error of ϵm=0.74\epsilon_m = 0.74 (Extended Data Figure 1,Extended Data Figure 1). Including the physics prior at the balanced weight λbalE\lambda^{E}_{\mathrm{bal}} accurately recovers all three components and reduces the error to ϵm=0.11\epsilon_m = 0.11.

Finally, because the material parameters supplied to the physics-based reconstruction are often only approximately known in experiment, we use the synthetic disk to test how sensitive the reconstructed state is to parameter mismatch. When we reconstruct the full dataset at λbalE\lambda^{E}_{\mathrm{bal}} while offsetting the input DD or MsM_s by ±30%\pm30\%, the error ϵm\epsilon_m increases from 0.110.11 to at most 0.150.15, whereas offsetting AA by +30%+30\% and −30%-30\% yields ϵm=0.11\epsilon_m = 0.11 and

Reconstructions from sparse and single-axis projection data

caption: Reconstructions from sparse and single-axis projection data. One row per acquisition condition: the full dual-axis measurement combining both tilt series (67 projections), a sparse dual-axis acquisition retaining only every eighth projection of each series (9), and the two single-axis acquisitions, xzxz series alone (29) and yzyz series alone (38). a, Acquisition in the geometry of Figure 1a: the beam directions swept by the xzxz series (blue) and by the yzyz series (orange), or the nine retained beam directions. b, Disk-plane mxm_x, mym_y, mzm_z cross sections, all reconstructed with the physics prior at λE=1015\lambda^{E}=10^{15}. Red frames mark the component that the single series does not measure: mym_y for the xzxz series and mxm_x for the yzyz series. Scale bar, 200 nm.

0.23, respectively (Extended Data Figure 1). The error rises most when AA is underestimated, because a weaker exchange stiffness smooths the measurement noise less. Apart from this case, parameter offsets of ±30%\pm30\% leave the reconstructed state largely unchanged, so normal uncertainty in the material parameters of the sample is tolerable.

Reconstruction from sparse and single-axis data

Time-resolved and field-dependent studies require a complete dual-axis tilt series for every time step or field value. Tomograms at successive time steps have been recorded [11], but faster acquisitions would allow for better temporal resolution and a wider range of phenomena and experimental parameters to be explored. To maximize the efficiency of these measurements, we test whether the physics prior can recover the magnetization from reduced projection data, either by sparse angular sampling or by omitting an entire tilt series. We reconstruct the disk under four acquisition conditions (Figure 3): the full dual-axis measurement, a sparse dataset with only every eighth projection of each tilt series (9 projections), and either tilt series alone (xzxz series or yzyz series). All four use λE=1015\lambda^{E}=10^{15}, ten times the balanced weight λbalE\lambda^{E}_{\mathrm{bal}}, to compensate for the reduced number of projections. Both the sparse and the single-axis reconstructions reproduce the chiral stripe pattern of the full dataset (Figure 3b). A single tilt series rotates the sample about one axis, so the magnetization component along that axis never aligns with the beam and goes unmeasured (§ 0.1). The physics prior fills it in: the micromagnetic energy couples the unmeasured component to the two measured ones, so all three come out consistent (red-framed panels in Figure 3b). The two single-axis reconstructions differ mainly at the edge of the sample, where the disk mask fits worst, and noise differences between the two series slightly change the stripe pattern (Extended Data Figure 2).

Figure 3. Sensitivity of the DD selection to the energy weight λE\lambda^{E} and the magnetization scale γ\gamma. Data misfit LdataL_{\mathrm{data}} versus DD, as in the DD selection of Figure 4. a, Curves for a sweep of λE\lambda^{E} at the default γ\gamma, colored by log⁡10λE\log_{10}\lambda^{E}. The dotted line marks the literature D0D_{0}. At small λE\lambda^{E} the data term predominately shapes m\mathbf{m} and no dip forms. The dip sharpens through the intermediate regime and freezes past the knee (λE≳1017\lambda^{E} \gtrsim10^{17}), where m\mathbf{m} settles into the energy minimum nearest to the data and the curves for the highest decades coincide. The selection weight λhighE=1020\lambda^{E}_{\mathrm{high}} = 10^{20} lies in this regime. There the minimum stays at D=0.66×10−3 J m−2D = 0.66 \times10^{-3}\ \mathrm{J\,m}^{-2} on the 0.03×10−3 J m−20.03 \times10^{-3}\ \mathrm{J\,m}^{-2} grid of this sweep, the value read off the refined grid of Figure 4a. b, The same curves at λhighE\lambda^{E}_{\mathrm{high}} for γ\gamma between 1×10−41 \times10^{-4} and 6×10−46 \times10^{-4}, with D0D_{0} dotted as in a. A larger γ\gamma deepens the dip and a smaller γ\gamma flattens it. The minimum stays at D=0.66×10−3 J m−2D = 0.66 \times10^{-3}\ \mathrm{J\,m}^{-2} for every γ\gamma.

However, agreement between the reduced and full reconstructions of the measured disk does not show that the unmeasured component is correct. We therefore repeat the four acquisitions on the synthetic disk, whose magnetization is known. At the weight of Figure 3 (λE=1015\lambda^{E}=10^{15}), the reconstruction recovers the three-dimensional helical winding in all three components for every acquisition, with εm\varepsilon_m between 0.08 and 0.16 (Extended Data Figure 1), compared with 0.74 for the complete dataset without the physics prior. Consequently, a small fraction of the projections, or a single tilt series alone, suffices to recover the three-dimensional texture, so that vector tomography experiments become faster and can cover more experimental conditions.

Reading material parameters from the data misfit

The magnetic texture of the disk forms through the competition between the micromagnetic interactions, so the measured texture should contain information about the material parameters. Here, we explore whether the tomographic reconstruction itself can determine the material parameters of the sample. However, varying the assumed parameters has little effect at the balanced weight λbalE\lambda^{E}_{\mathrm{bal}}, where the reconstruction is primarily guided by the measurement data (Sec. II). To increase the sensitivity of the reconstruction to the material parameters, we instead raise λE\lambda^{E} to a high weight (λhighE\lambda^{E}_{\mathrm{high}} in Figure 2b). In this regime, the physics prior dominates the objective function, forcing m\mathbf{m} toward a low-energy state dictated by the assumed parameters. If those parameters differ from the sample’s true constants, the resulting texture does not match the measured projections and the data misfit Ldata\mathcal{L}_{\mathrm{data}} rises. Consequently, we can find the best-matching parameters by reconstructing across candidate values at λhighE\lambda^{E}_{\mathrm{high}} and looking for the minimum in Ldata\mathcal{L}_{\mathrm{data}}.

We first vary one parameter at a time around the literature values A0=5.73 pJ m−1A_{0}=5.73\,\mathrm{pJ}\,\mathrm{m}^{-1}, D0=0.60×10−3 J m−2D_{0}=0.60 \times10^{-3}\,\mathrm{J}\,\mathrm{m}^{-2}, and Ms=461 kA m−1M_{s}=461\,\mathrm{kA}\,\mathrm{m}^{-1} [20], performing a separate tomographic reconstruction of the disk at λE=1020\lambda^{E}=10^{20} for each value (Figure 4a). Across a ±30%\pm30\% range in MsM_{s}, Ldata\mathcal{L}_{\mathrm{data}} changes by less than 1%, showing that the data misfit is relatively insensitive to the saturation magnetization. In contrast, AA and DD form two wells of equal depth with minima at 0.91A00.91A_{0} and 1.10D01.10D_{0}. Because a 9% decrease in AA and a 10% increase in DD change the ratio A/DA/D by the same factor, these mirrored minima indicate that we are only sensitive to the ratio of A/DA/D. To test this across

Figure 4. Sensitivity of the data misfit to the material parameters. a, Data misfit Ldata\mathcal{L}_{\mathrm{data}} at λE=1020\lambda^{E}=10^{20} when one parameter is varied, with the other two fixed at the literature values A0A_{0}, D0D_{0}, MsM_{s} [20]. The misfit stays flat in MsM_{s}, whereas AA and DD trace two wells with minima on either side of the literature point, 9% below A0A_{0} and 10% above D0D_{0} (dotted lines, the vertices of the dashed parabolas). b, Data misfit over a 13×1313 \times13 grid in AA and DD (deviation from A0A_{0}, D0D_{0}) at λE=1020\lambda^{E}=10^{20}. The misfit is low along a valley of constant A/DA/D. The white line is the fitted trough of that valley, the red and blue lines are the A=A0A=A_{0} and D=D0D=D_{0} cuts of a. c, Depth-averaged mzm_{z} of the four reconstructions, with AA and DD labeled relative to A0A_{0}, D0D_{0}. Along the valley: the minimum of the A=A0A=A_{0} cut (orange) and the on-valley cell farthest from it (brown). Across the valley: the low-energy states at A=A0A=A_{0} on the grid edges, with D=0.70D0D=0.70D_{0} and 1.30D01.30D_{0} (cyan and pink). Markers of the same colors mark these reconstructions in b, and those at A=A0A=A_{0} in a. Black marks the literature values. d, Mean of the four measured XMCD projections within 2∘2^\circ of normal incidence (θxz=±1∘\theta_{xz}=\pm1^\circ and θyz=±1∘\theta_{yz}=\pm1^\circ), for comparison with the textures in c. Scale bars, 200 nm.

both parameters simultaneously, we reconstruct the disk over a two-dimensional (A,D)(A,D) grid (Figure 4b). The misfit forms a continuous valley along a line of constant A/DA/D, confirming that the tomographic data fix only the relative strength of exchange and DMI. The constant-ratio valley reflects the competing roles of exchange and DMI in the micromagnetic energy, where A/DA/D sets the equilibrium winding period of the helix (given by the pitch LD=4πA/DL_{D}=4\pi A/D in an unbounded crystal) [34]. Along the valley floor, parameter pairs with widely different absolute values therefore converge to nearly identical stripe textures with the same period (Figure 4c, orange and brown). In contrast, when we step off the valley by changing DD by ∓30%\mp30\% at fixed A=A0A=A_0, the stripe period shifts noticeably: D=0.42×10−3 J m−2D=0.42\times10^{-3}\ \mathrm{J\,m^{-2}} (cyan) widens the stripes, whereas D=0.78×10−3 J m−2D=0.78\times10^{-3}\ \mathrm{J\,m^{-2}} (pink) tightens the winding. Only parameter pairs along the valley floor reproduce the stripe period seen in the measured face-on projections (Figure 4d).

With AA fixed at A0A_0, the minimum at 1.10D01.10D_0 gives D=0.66×10−3 J m−2D=0.66\times10^{-3}\ \mathrm{J\,m^{-2}} (Figure 4a). This well is shallow: Ldata\mathcal{L}_{\mathrm{data}} rises by less than 1% within 0.03×10−3 J m−20.03\times10^{-3}\ \mathrm{J\,m^{-2}} of the optimum. Even so, its minimum stays at the same DD for every energy weight from λE=1016\lambda^{E}=10^{16} to 102010^{20} (Extended Data Fig. 3a). Only at lower weights does the well flatten and shift toward smaller DD, before disappearing near the balanced weight λbalE=1014\lambda^{E}_{\mathrm{bal}}=10^{14}.

Our readout lies 10% above the literature value D0D_0, although our disk winds with the same 120 nm period (Figure 4d) as the nanowire from which D0D_0 was derived [20]. The literature value follows from inserting this period into the bulk relation D=4πA0/LD=4\pi A_0/L, which describes a helix in an unbounded crystal, whereas our readout models the full micromagnetic energy of the confined disk. We test this difference with the synthetic disk, whose DD is known: relaxed at A=A0A=A_0 and D=0.66×10−3 J m−2D=0.66\times10^{-3}\ \mathrm{J\,m^{-2}} (Methods), it forms stripes with a 122 nm period, longer than the 109 nm that the bulk relation predicts. Applied to this period, the bulk relation underestimates DD by about 10%, as in the measured disk. Our readout, in contrast, returns the DD used to relax the synthetic disk from its simulated projections (Extended Data Figure 1).

Even at the optimal DD, the high-weight reconstruction (λhighE\lambda^{E}_{\mathrm{high}}) has a larger data misfit than the balanced reconstruction (Ldata=0.93\mathcal{L}_{\mathrm{data}}=0.93 versus 0.83, Figure 2a). In a real sample, however, pinning sites, local defects, or focused-ion-beam damage can distort the texture in ways that our energy model does not capture, increasing the data misfit at λhighE\lambda^{E}_{\mathrm{high}}. In our disk these local distortions are small enough that the overall texture still lies close to a minimum of Eq. (4), allowing a clear dip in Ldata\mathcal{L}_{\mathrm{data}}. If strong pinning instead traps a sample in a metastable state that Eq. (4) cannot reproduce, the high-weight reconstruction converges to a state far from the measured texture for every parameter set, leaving Ldata\mathcal{L}_{\mathrm{data}} uniformly high and preventing a readout.

Conclusions

We have presented an approach to imaging nanoscale magnetic textures that incorporates the micromagnetic energy into the objective function, with a single weight λE\lambda^{E} selecting between two operating regimes of the same reconstruction. In the balanced regime the prior helps recover the texture from noisy, incomplete measurements, and the result tolerates the typical uncertainty of the assumed material parameters. A larger weight on the micromagnetic energy switches to the second regime, in

which the data misfit can be used to read out the parameters that shape the texture. Moreover, a parameter’s impact on the reconstruction matches its sensitivity in the readout: parameters that barely affect the texture cannot be extracted, but they also barely change the reconstructed state, whereas parameters that strongly shape the texture can be determined directly from the misfit. In the helimagnetic Co8_8Zn9_9Mn3_3 disk, the reconstruction is relatively insensitive to MsM_s, whereas the ratio A/DA/D sets the period of the helical winding. At the literature exchange stiffness A0A_0, the readout returns D=0.66×10−3 J m−2D=0.66\times10^{-3}\ \mathrm{J\,m^{-2}}, 10% above D0D_0, although the disk winds with the 120 nm period from which D0D_0 was derived. The bulk relation behind D0D_0 describes an unbounded crystal and underestimates DD in the confined disk. The ratio A/DA/D therefore has to be measured in the nanostructure itself.

The sparse-data recovery and the parameter readout together suggest a strategy for experimental design. Dense dual-axis tilt series, typically required for high-resolution three-dimensional imaging, take a significant amount of beamtime in vector tomography. With the physics prior, far fewer projections are required, and even a single tilt axis can provide a robust reconstruction. Every projection a reconstruction can spare buys beamtime for additional specimens, magnetic fields, or time points within one allocation, and thus for more complex experiments such as in situ studies. A single well-sampled measurement of a prepared reference state, such as a helical texture whose period pins the constants, could then validate the energy model and fix the material parameters, while subsequent conditions are imaged with sparse acquisitions.

We demonstrate this physics-informed approach with XMCD tomography, but automatic differentiation in micromagnetic software [28, 35] allows the physics prior to be paired with any differentiable forward model and extended with further interactions of the host material. The same approach can therefore be extended to other magnetic imaging modalities, including electron holographic and Lorentz tomography [12, 13] and laminography [11].

Methods

Data acquisition and preprocessing

a. Sample. The Co8_8Zn9_9Mn3_3 disk of Extended Data Table I sits on a focused-ion-beam support wedge. The disk was cut from a single crystal grown at the University of Warwick by the Bridgman method [14]. The table gives its acquisition and discretization. The material parameters MsM_s, A0A_0, and D0D_0 are those of Ref. [20] for the same compound: MsM_s from demagnetization data, and A0A_0 with D0D_0 chosen there to reproduce the helical period LD=4πA0/D0=120 nmL_D=4\pi A_0/D_0=120\ \mathrm{nm} imaged in a 200 nm thick nanowire. b. X-ray data acquisition. We acquired the projections at room temperature using dichroic soft X-ray ptychography [8, 23] on the I08-1 ptychography end station of Diamond Light Source under proposal MG37535-2. An APPLE-II undulator produced circularly polarized X-rays of either helicity, set by its gap and phase, and the Au 600 lines/mm grating of the SX-700 plane-grating monochromator [36] tuned them 3 eV below the Co L3_3 edge (778.1 eV). A 10 µm pinhole defined the secondary source, and a Fresnel zone plate with a diameter of 333 µm, an outermost zone width of 70 nm, and a mean focal length of 14.156 mm between 770 and 807 eV focused the beam. The sample sat on an L-bracket on an SR-5714C rotation stage (SmarAct GmbH), about 50 µm downstream of the focus, and we scanned it through the beam along a spiral path to avoid grid artifacts. An in-vacuum Axis-SXRF-2020EUV sCMOS detector (Axis Photonique Inc.) recorded the diffraction patterns. We measured the two polarizations separately as C+C^{+} and C−C^{-}, each yielding a series of diffraction patterns. At the absorption peak, resonant attenuation would block transmission through the disk, whereas at this pre-edge energy the transmission remains high while the magnetic phase contrast persists [14]. We reconstructed the complex-valued object projections (amplitude and phase contrast images) and the probe with PtyPy [37] after cropping the diffraction patterns to 1024 × 1024 pixels and binning them to 512 × 512 pixels, running 1000 iterations of the relaxed averaged alternating reflections (RAAR) engine, 200 iterations of maximum-likelihood refinement, and another 500 RAAR iterations.

c. Preprocessing. Each complex projection carries an absorptive channel (the log amplitude) and a dispersive channel (the phase). In either channel, we combine the projections of the two polarizations to isolate the electronic term cc and the XMCD term yy,

c=−12(C++C−),y=12(C−−C+).(5)c=-\frac{1}{2}\left(C^{+}+C^{-}\right), \qquad y=\frac{1}{2}\left(C^{-}-C^{+}\right). \tag*{(5)}

Both encode the same projected moment, but at the pre-edge energy the phase XMCD has the higher signal-to-noise ratio, and therefore is selected as the reconstruction input. Starting from a coarse pre-alignment of the projections, we fit an ellipse to the projected disk outline in every frame and shift its center onto the rotation axis [38], register the exposures of each angle (both polarizations and any repeats) against each other on the electronic channel before averaging them, and apply the composed shift in a single cubic interpolation per exposure. We then remove the phase ramp of every projection by fitting a plane to the phase over the air region, detected by an Otsu threshold [39] on the transmission amplitude, and subtracting it. For the electronic channel, we subtract from cc its median outside the disk support in every frame, which normalizes the transmitted amplitude to the background.

d. Tilt geometry. The stage angles enter the forward model through one rotation per series about its axis. We calibrate the zero point θ0\theta_0 of each series from the same disk outlines. The disk outline is a circle when the disk faces the beam and an ellipse that narrows as the disk tilts away. We fit the aspect ratio of this ellipse in all frames of a series at once with a(θ)=s[cos⁡(θ+θ0)+k∣sin⁡(θ+θ0)∣]a(\theta)=s[\cos(\theta+\theta_0)+k|\sin(\theta+\theta_0)|], where θ\theta is the stage angle recorded with each frame in the raw data. The fit has three free parameters: the zero point θ0\theta_0, a scale ss, and kk, which accounts for the blurred, partly absorbing rim. We apply θ0\theta_0 rounded to the nearest degree, the precision of the stage labels (Extended Data Table I).

Magnetic reconstruction framework

e. Loss functional. Both the physics prior and the Tikhonov reconstruction minimize a squared data misfit plus one regularizer RR with weight λ\lambda,

L(m)=Ldata2(m)+λR(m).(6)\mathcal{L}(\mathbf{m})=\mathcal{L}_{\mathrm{data}}^{2}(\mathbf{m})+\lambda R(\mathbf{m}). \tag*{(6)}

In the data term of Eq. (3), pi\mathbf{p}_i is the forward model of Eq. (7), yi\mathbf{y}_i the measured XMCD projection, and the norms run over the projection pixels of projection ii. Each residual is normalized by that projection’s own measurement power, so every view carries equal weight. The residual is always evaluated on the full projection frame: the disk mask acts on the magnetization field, never on the data residual.

f. Forward model. The Tikhonov reconstructions use an isotropic 120 × 120 × 120 tomography grid with a cell spacing of ≈ 9.96 nm (one cell per projection pixel across the central projection window). The physics reconstructions instead run on a box that just encloses the disk (Extended Data Table I). In the lab frame the beam runs along z^\hat{\mathbf{z}}, the projection pixel (x,y)(x,y) spans the transverse plane, and the stage rotation Ri\mathbf{R}_i carries a sample point r\mathbf{r} to the lab position Rir\mathbf{R}_i\mathbf{r} (Figure 1a). Equation (1) written out with the magnetization scale γ\gamma and the disk mask ww gives the value pi(x,y)p_i(x,y) of the projection pi(m)\mathbf{p}_i(\mathbf{m}) at projection pixel (x,y)(x,y) as the line integral along the beam,

pi(x,y)=γ∫[wm⋅k^i](RiTr) dz,r=(x,y,z).(7)p_i(x,y)=\gamma\int\left[w\mathbf{m}\cdot\hat{\mathbf{k}}_i\right](\mathbf{R}_i^{T}\mathbf{r})\,\mathrm{d}z,\qquad\mathbf{r}=(x,y,z). \tag*{(7)}

where the sample-frame field is evaluated at the sample point the rotation carries to r\mathbf{r}, and γ\gamma is the amplitude the projector sees at every cell inside the mask: under the unit-norm constraint ∣m∣=1|\mathbf{m}|=1 the projected magnetization has ∣γm∣=γ|\gamma\mathbf{m}|=\gamma everywhere in the material. It collects MsM_s, the magnetic scattering factor, the atomic density and the cell size into one constant [8], which we read from the Tikhonov reconstructions (‘Magnetization-scale calibration’ below). We evaluate this integral on the cell grid with the differentiable projector of SCICO [24].

g. Tikhonov regularization. The Tikhonov reconstruction treats m\mathbf{m} as an unconstrained Cartesian vector field inside the disk mask ww, without a magnetization norm constraint: it replaces m\mathbf{m} by wmw\mathbf{m} in the forward model, so cells outside the mask carry no magnetization and the penalty R(m)=∥m∥2\mathcal{R}(\mathbf{m}) = \lVert\mathbf{m}\rVert^2 of Eq. (6) sums ∣m∣2\lvert\mathbf{m}\rvert^2 over the disk. We minimize the resulting objective with the L-BFGS implementation in JAXopt [40], relying on its zoom line search to set the step size. On the measured data we use λbalTik=1\lambda_{\mathrm{bal}}^{\mathrm{Tik}} = 1 (Extended Data Figure 4). An unmasked variant reconstructs the field freely on the full box.

Extended Data Figure 4. Tikhonov reconstructions and the magnetization scale γ\gamma. a, L-curve of the Tikhonov sweep, solution norm ∥m∥\lVert\mathbf{m}\rVert versus the data misfit Ldata\mathcal{L}_{\mathrm{data}}, for the unmasked field on the full box and for the mask-confined field, with λbal=1\lambda_{\mathrm{bal}}=1 marked. b, The learned magnetization scale γ⋆\gamma^\star (interior median ∣m∣\lvert\mathbf{m}\rvert) versus λTik\lambda^{\mathrm{Tik}} for the same two sweeps, λbal\lambda_{\mathrm{bal}} marked as in a. γ⋆\gamma^\star changes monotonically with λTik\lambda^{\mathrm{Tik}}. The fixed γ\gamma of the physics reconstructions (dashed line, Methods, ’Magnetic reconstruction framework’) is the mask-confined value at λbalTik\lambda_{\mathrm{bal}}^{\mathrm{Tik}}. c, The unmasked reconstruction at λbalTik\lambda_{\mathrm{bal}}^{\mathrm{Tik}}, components mxm_x, mym_y, mzm_z in the disk plane (xyxy, left) and a side view (yzyz, right), shown in full with the mask boundary drawn gray dashed. The field smears across the full depth and leaks outside the mask. d, The mask-confined reconstruction at λbalTik\lambda_{\mathrm{bal}}^{\mathrm{Tik}}, in the same layout. The masked volume carries the same in-plane stripe pattern at higher contrast, and the amplitude rises from γ⋆=2.5×10−4\gamma^\star= 2.5 \times10^{-4} (unmasked) to 4.1×10−44.1 \times10^{-4}. Each reconstruction is shown on one color scale, set by its 99.5th percentile over all components. Scale bars, 200 nm.

h. Micromagnetic energy regularization. The physics reconstruction sets R=EtotalR = E_{\mathrm{total}} and λ=λE\lambda= \lambda^{E}, the plain micromagnetic energy of Eq. (4) in joules, so λE\lambda^{E} carries the inverse units of energy. The unit-norm constraint ∣m∣=1\lvert\mathbf{m}\rvert= 1 needs the mask ww to separate material from vacuum and the scale γ\gamma of Eq. (7) to match the measured amplitude (Extended Data Table I gives the bounding box). With exchange stiffness AA, saturation magnetization MsM_{\mathrm{s}}, demagnetizing field Hdem\mathbf{H}^{\mathrm{dem}}, and bulk Dzyaloshinskii–Moriya constant DD, its terms take their standard form [19],

Eex=∫ΩA∣∇m∣2 dV,E^{\mathrm{ex}} = \int_{\Omega} A \lvert\nabla\mathbf{m}\rvert^2 \,\mathrm{d}V,
Edem=−12μ0Ms∫Ωm⋅Hdem dV,E^{\mathrm{dem}} = -\frac{1}{2}\mu_0 M_{\mathrm{s}} \int_{\Omega} \mathbf{m}\cdot\mathbf{H}^{\mathrm{dem}} \,\mathrm{d}V,
Edmi=∫ΩDm⋅(∇×m) dV(8)E^{\mathrm{dmi}} = \int_{\Omega} D\mathbf{m}\cdot(\nabla\times\mathbf{m})\,\mathrm{d}V \tag*{(8)}

We compute the energy with NeuralMag [28] on its JAX backend. It evaluates the local terms by finite differences and the demagnetizing field by FFT convolution.

The energy must be evaluated on a grid that resolves the characteristic length scales of the material, the exchange length ℓex=2Aμ0Ms2≈6.6 nm\ell_{\mathrm{ex}} = \sqrt{\frac{2A}{\mu_0 M_{\mathrm{s}}^2}} \approx6.6\ \mathrm{nm} and the DMI length ℓD=2A/D≈17.4 nm\ell_D = 2A/D \approx17.4\ \mathrm{nm}. The tomography cell of ≈9.96 nm\approx9.96\ \mathrm{nm}, the projection pixel, bounds the spatial resolution and samples the 120 nm120\ \mathrm{nm} helical period with ≈12\approx12 cells, but it sits above ℓex\ell_{\mathrm{ex}}, so we refine it by a factor of two and evaluate the energy on the resulting 4.98 nm4.98\ \mathrm{nm} physics grid (≈24\approx24 cells per period), below both micromagnetic length scales. We reconstruct the magnetization on this finer grid and block-average it onto the tomography grid before projection.

i. Optimization details. Riemannian Adam [41], the manifold generalization of Adam [42], keeps ∣m∣=1\lvert\mathbf{m}\rvert= 1 exactly at every step (custom JAX implementation). We run at a constant learning rate of 0.40.4 with the standard coefficients (β1=0.9\beta_1 = 0.9, β2=0.999\beta_2 = 0.999, ϵ=10−8\epsilon= 10^{-8}) for 300 to 400 iterations from a cold start at m=0\mathbf{m} = 0. The zero vector lies outside the unit sphere, so the first step is a Euclidean gradient step whose normalization places every cell on the sphere, and all later steps are Riemannian. Each physics reconstruction fits on one 24 GB24\ \mathrm{GB} MIG slice of an NVIDIA RTX PRO 6000 Blackwell GPU in single precision. The energy, the projector, and the optimizer update fuse into one kernel under just-in-time compilation [43], which costs about 55 s55\ \mathrm{s} once per configuration. On the full GPU, a compiled 400-iteration reconstruction then takes 13 s13\ \mathrm{s}, and λ\lambda enters the kernel as a runtime argument, so the nine-point L-curve sweep (λE=1012\lambda^{E} = 10^{12} to 102010^{20} in decade steps) reuses one compilation and completes in about three minutes.

j. Disk mask. We obtain the in-plane disk outline by Otsu thresholding [39] the electronic channel of the zero-tilt projection (Figure 1b). The outline includes the short neck that joins the disk to the support wedge, so the reconstructed texture continues into it. We set the mask to this uniform outline across 30 cells in the zz-direction, matching the fabricated 300 nm sample thickness (Extended Data Table I). A separate scalar tomography of the electronic channel confirms this mask (Extended Data Figure 5).

Extended Data Figure 5. Validation of the disk mask through total-variation (TV) tomography of the electronic channel. a, L-curve: total variation LTVL_{\mathrm{TV}} versus the electronic-channel data misfit LdataabsL_{\mathrm{data}}^{\mathrm{abs}} as the TV weight λTV\lambda^{\mathrm{TV}} is swept. The misfit is flat up to the operating point λTV=1\lambda^{\mathrm{TV}}=1 (red) and rises beyond it, where the reconstruction over-smooths. b, The disk mask used in all magnetic reconstructions, shown in three dimensions with the three slice planes of c: the disk plane (xyxy) at mid-slab and the two side views (xzxz, yzyz) through the disk center. c, Reconstructed density at λTV=0.01\lambda^{\mathrm{TV}}=0.01, 1 (operating point), and 100, shown in the disk plane (xyxy, top row) and the two side views (xzxz, yzyz). At λTV=0.01\lambda^{\mathrm{TV}}=0.01 the density is streaky, and at λTV=100\lambda^{\mathrm{TV}}=100 it is over-smoothed. The density extends beyond the slab in zz because of the missing wedge, more so at larger λTV\lambda^{\mathrm{TV}}. Scale bar, 200 nm.

k. Electronic-channel tomography. We validate the disk mask with a scalar tomography of the electronic channel cc of Eq. (5), taken from the log amplitude. We reconstruct a non-negative density ρ\rho on a free 1203120^3 box of the tomography grid, without the mask, by minimizing

L(ρ)=(Ldataabs)2(ρ)+λTVLTV(ρ),(9)\mathcal{L}(\rho) = \left(\mathcal{L}_{\mathrm{data}}^{\mathrm{abs}}\right)^2(\rho) + \lambda^{\mathrm{TV}}\mathcal{L}_{\mathrm{TV}}(\rho), \tag*{(9)}

where the squared misfit averages the relative residual over all NN projections ci\mathbf{c}_i,

(Ldataabs)2(ρ)=1N∑i=1N∥Riρ−ci∥2∥ci∥2,(10)\left(\mathcal{L}_{\mathrm{data}}^{\mathrm{abs}}\right)^2(\rho) = \frac{1}{N}\sum_{i=1}^{N}\frac{\lVert\mathcal{R}_i\rho-\mathbf{c}_i\rVert^2}{\lVert\mathbf{c}_i\rVert^2}, \tag*{(10)}

the electronic-channel counterpart of Eq. (3). Here Ri\mathcal{R}_i is the scalar line integral of Eq. (7) along the beam of projection ii, and LTV\mathcal{L}_{\mathrm{TV}} is the anisotropic total variation, the absolute differences between neighboring cells averaged over the cells of the box. Total variation favors piecewise-constant densities with sharp edges, the form of a uniform disk in air. We enforce ρ≥0\rho\ge0 through ρ=u2\rho= u^2 and run 300 L-BFGS iterations from a uniform start for each λTV\lambda^{\mathrm{TV}} from 10−210^{-2} to 10210^2 in decade steps. The root Ldataabs\mathcal{L}_{\mathrm{data}}^{\mathrm{abs}} is the misfit axis of the L-curve (Extended Data Figure 5).

l. Magnetization-scale calibration. The magnetization scale γ\gamma of Eq. (7) matches the unit-norm magnetization to the measured XMCD amplitude [8]. We read it from the Tikhonov reconstructions, whose free amplitude defines a learned scale γ⋆(λTik)\gamma^\star(\lambda^{\mathrm{Tik}}) as the interior median ∣m∣\lvert\mathbf{m}\rvert. At λbalTik=1\lambda_{\mathrm{bal}}^{\mathrm{Tik}} = 1, the mask-confined Tikhonov reconstruction and its unmasked variant give γ⋆≈4.1×10−4\gamma^\star\approx4.1\times10^{-4} and 2.5×10−42.5\times10^{-4}. All physics reconstructions use γ=4×10−4\gamma= 4\times10^{-4}, the mask-confined value (Extended Data Figure 4).

Synthetic validation

m. Ground truth generation. The synthetic ground truth is a disk of the measured sample’s dimensions (diameter 1.1 μm1.1\ \mu\mathrm{m}, thickness 300 nm300\ \mathrm{nm}) initialized from the measurement itself: we take the reconstruction of the full measurement without the energy term (λE=0\lambda^{E} = 0), a unit field on the 4.98 nm4.98\ \mathrm{nm} physics grid (Extended Data Fig. 1a). We relax this state with the Landau–Lifshitz–Gilbert equation [32, 33] in NeuralMag [28] to equilibrium at the Extended Data Table I parameters with D=0.66×10−3 J m−2D = 0.66 \times10^{-3}\ \mathrm{J\,m}^{-2}, and block-average it onto the tomography grid. The relaxation turns the noisy seed into a smooth labyrinth of stripes that follows the arrangement of the measured texture (Extended Data Fig. 1a).

n. Measurement simulation. We project this state with the forward model of Eq. (7), the operator the reconstruction inverts, at the 29+3829 + 38 calibrated stage angles of the measurement and at the magnetization scale γ\gamma of the reconstructions, so the measured noise amplitude carries over directly. The noise is a stationary Gaussian field with the radially averaged power spectrum of air-only patches of the measured XMCD channel, scaled to the median standard deviation of the measured air pixels, and added with opposite sign to the two polarizations.

o. Evaluation. We reconstruct the synthetic projections using the same solver settings and disk mask as the experimental data. Fidelity is quantified by the relative reconstruction error εm=∥m⋆−mGT∥/∥mGT∥\varepsilon_m = \lVert\mathbf{m}^{\star} - \mathbf{m}^{\mathrm{GT}} \rVert/ \lVert\mathbf{m}^{\mathrm{GT}} \rVert, evaluated inside the disk mask.

Use of large language models

We used the large language model tools Claude (Anthropic) and Antigravity (Google) to assist with writing the analysis and reconstruction code, with the data analysis, and with drafting and editing the manuscript text. The authors reviewed, tested, and revised all output and take full responsibility for the code, the results, and the content of this article.

Data and code availability

The measured XMCD projections, the preprocessed solver inputs, the synthetic ground truth, and every reconstruction shown in the main text and the Extended Data will be archived at Zenodo upon journal publication. The preprocessing and reconstruction code (the xmcdrecon package, JAX and NeuralMag), with the recipes and commands behind every measured-data reconstruction of this work, will be released on GitHub and archived at Zenodo upon journal publication.

Acknowledgments

We acknowledge Diamond Light Source for time on instrument I08-1 under proposal MG37535-2. This research was funded in whole or in part by the Austrian Science Fund (FWF) 10.55776/PIN1629824 and 10.55776/P34671. P.M., R.Y., M.R.F., and C.D. acknowledge funding from the Max Planck Society Lise Meitner Excellence Program and funding from the European Research Council (ERC) under the ERC Starting Grant 3DNANOQUANT 101116043. A.F.-P. acknowledges support by the European Community under the Horizon 2020 Program, Contract No. 101001290 (3DNANO¬MAG). The crystal growth at Warwick was supported by two Engineering and Physical Sciences Research Council grants: EP/Z535874/1, “Single Crystal Growth at Warwick”, and EP/Z534067/1, “Controlling the Formation and Properties of Topological Magnetic Phenomena”. For the purpose of open access, the authors have applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission.

Author contributions

A.S. developed the physics-informed reconstruction framework, carried out the reconstructions and the analysis, and wrote the manuscript. P.M. prepared the samples, performed the X-ray ptychographic tomography measurements, the ptychographic reconstruction and coarse alignment of the projections, and contributed to the magnetic reconstructions and the analysis. L.J.R.H. and B.K. set up and operated beamline I08-1 and supported the ptychographic tomography measurements. L.T., R.Y., and M.R.F. contributed to the X-ray measurements. D.A.M. and G.B. grew the single crystals. A.C. contributed to the implementation of the reconstruction method. D.S. contributed to the supervision of the project. A.F.-P. and M.G.-S. contributed to the conception and planning of the project and to writing the proposal. C.D. supervised the experimental work. C.A. is the lead developer of the underlying NeuralMag micromagnetic framework and contributed to and supervised the development of the reconstruction method. All authors discussed the reconstructions, the analysis, and the results, and commented on the manuscript.

Competing interests

The authors declare no competing interests.

Extended Data

experimental▪ synthetic
Saturation magnetization MsM_s460.55 kA m−1^{-1}
Exchange stiffness A0A_05.729 pJ m−1^{-1}
Bulk DMI constant D0D_0 (literature)0.60 × 10−3^{-3} J m−2^{-2}
Bulk DMI constant DD (readout, used in all reconstructions)0.66 × 10−3^{-3} J m−2^{-2}0.66 × 10−3^{-3} J m−2^{-2} (truth)
Tomography cell size≈ 9.96 nm
Physics cell size≈ 4.98 nm (2× refine)
Stored projection window250 × 250 px
Thickness300 nm
Diameter1.1 µm
xzxz series (projections, range)29, [−41, +15]°
yzyz series (projections, range)38, [−30, +47]°
Projectionsmeasuredsimulated at the measured angles with measured noise
Angle offset (xzxz series, yzyz series)15°, 33°—
Reconstruction box (tomography cells)120 × 120 × 30
Physics grid (cells)240 × 240 × 60
Face-on stripe period (nm)120122
Main figures1–4—
Extended Data figures2–51

Table I. The two datasets of this work, with their material parameters, acquisition, and discretization. The upper block lists the literature parameters of Ref. [20], the DMI constants, and the two grids. The lower block lists the acquisition and geometry of each dataset, including the zero-point offset added to the logged stage angles of each measured series (Methods, ’Data acquisition and preprocessing’). The synthetic dataset is described in Methods, ’Synthetic validation’. The last two rows list the figures that use each dataset.

References

  1. [1]S. Seki, M. Suzuki, M. Ishibashi, R. Takagi, N. D. Khanh, Y. Shiota, K. Shibata, W. Koshibae, Y. Tokura, and T. Ono, Direct visualization of the three-dimensional shape of skyrmion strings in a noncentrosymmetric mag- net, Nature Materials 21, 181 (2022).
  2. [2]M. T. Birch, D. Cortés-Ortuño, L. A. Turnbull, M. N. Wilson, F. Groß, N. Träger, A. Laurenson, N. Bukin, S. H. Moody, M. Weigand, G. Schütz, H. Popescu, R. Fan, P. Steadman, J. a. T. Verezhak, G. Balakrishnan, J. C. Loudon, A. C. Twitchett-Harrison, O. Hovorka, H. Fanghor, F. Y. Ogrin, J. Gräfe, and P. D. Hatton, Real-space imaging of confined magnetic skyrmion tubes, Nature Communications 11, 1726 (2020).DOI
  3. [3]N. Kent, N. Reynolds, D. Raftrey, I. T. G. Campbell, S. Virasamy, S. Dhuey, R. V. Chopdekar, A. Hierro-Rodriguez, A. Sorrentino, E. Pereiro, S. Ferrer, F. Hellman, P. Sutcliffe, and P. Fischer, Creation and observation of Hopfions in magnetic multilayer systems, Nature Communications 12, 1562 (2021).DOI
  4. [4]C. Donnelly, K. L. Metlov, V. Scagnoli, M. Guizar-Sicairos, M. Holler, N. S. Bingham, J. Raabe, L. J. Heyderman, N. R. Cooper, and S. Gliga, Experimental observation of vortex rings in a bulk magnet, Nature Physics 17, 316 (2021).DOI
  5. [5]N. Nagaosa and Y. Tokura, Topological properties and dynamics of magnetic skyrmions, Nature Nanotechnology 8, 899 (2013).DOI
  6. [6]A. Fert, N. Reyren, and V. Cros, Magnetic skyrmions: advances in physics and potential applications, Nature Reviews Materials 2, 17031 (2017).DOI
  7. [7]G. Schütz, W. Wagner, W. Wilhelm, P. Kienle, R. Zeller, R. Frahm, and G. Materlik, Absorption of circularly polarized x rays in iron, Physical Review Letters 58, 737 (1987).pubmed.ncbi.nlm.nih.gov/10035022
  8. [8]C. Donnelly, M. Guizar-Sicairos, V. Scagnoli, S. Gliga, M. Holler, J. Raabe, and L. J. Heyderman, Three-dimensional magnetization structures revealed with X-ray vector nanotomography, Nature 547, 328 (2017).DOI
  9. [9]C. Donnelly, S. Gliga, V. Scagnoli, M. Holler, J. Raabe, L. J. Heyderman, and M. Guizar-Sicairos, Tomographic reconstruction of a three-dimensional magnetization vector field, New Journal of Physics 20, 083009 (2018).DOI
  10. [10]A. Hierro-Rodriguez, C. Quirós, A. Sorrentino, L. M. Alvarez-Prado, J. I. Martín, J. M. Alameda, S. McVitie, E. Pereiro, M. Vélez, and S. Ferrer, Revealing 3D magnetization of thin films with soft X-ray tomography: magnetic singularities and topological charges, Nature Communications 11, 6382 (2020).DOI
  11. [11]C. Donnelly, S. Finizio, S. Gliga, M. Holler, A. Hrabec, M. Odstrčil, S. Mayr, V. Scagnoli, L. J. Heyderman, M. Guizar-Sicairos, and J. Raabe, Time-resolved imaging of three-dimensional nanoscale magnetization dynamics, Nature Nanotechnology 15, 356 (2020).DOI
  12. [12]D. Wolf, S. Schneider, U. K. Rößler, A. Kovács, M. Schmidt, R. E. Dunin-Borkowski, B. Büchner, B. Rellinghaus, and A. Lubk, Unveiling the three-dimensional magnetic texture of skyrmion tubes, Nature Nanotechnology 17, 250 (2022).DOI
  13. [13]C. Phatak, A. K. Petford-Long, and M. De Graef, Three-dimensional study of the vector potential of magnetic structures, Physical Review Letters 104, 253901 (2010).pubmed.ncbi.nlm.nih.gov/20867379
  14. [14]J. Neethirajan, B. J. Daurer, M. Di Pietro Martínez, A. Hrabec, L. Turnbull, Y. Yamamoto, M. R. Ferreira, A. Štefančič, D. A. Mayoh, G. Balakrishnan, Z. Pei, P. Xue, L. Chang, E. Ringe, R. Harrison, S. Valencia, M. Kazemian, B. Kaulich, and C. Donnelly, Soft X-Ray Phase Nanoscopy of Micrometer-Thick Magnets, Physical Review X 14, 031028 (2024).
  15. [15]R. A. Crowther, D. J. DeRosier, and A. Klug, The reconstruction of a three-dimensional structure from projections and its application to electron microscopy, Proceedings of the Royal Society of London. A Mathematical and Physical Sciences 317, 319 (1970).DOI
  16. [16]F. Bruckner, C. Abert, G. Wautischer, C. Huber, C. Vogler, M. Hinze, and D. Suess, Solving Large-Scale Inverse Magnetostatic Problems using the Adjoint Method, Scientific Reports 7, 40816 (2017).DOI
  17. [17]A. Silinga, A. Kovács, S. McVitie, R. E. Dunin-Borkowski, K. Fallon, and T. P. Almeida, Model-Based Iterative Reconstruction of Three-Dimensional Magnetisation in a Nanowire Structure Using Electron Holographic Vector Field Tomography, Microscopy and Microanalysis 31, ozaf043 (2025).DOI
  18. [18]W. Brown, Micromagnetics, Interscience tracts on physics and astronomy (Interscience Publishers, 1963).
  19. [19]C. Abert, Micromagnetics and spintronics: models and numerical methods, The European Physical Journal B 92, 120 (2019).DOI
  20. [20]M. T. Birch, D. Cortés-Ortuño, K. Litzius, S. Wintz, F. Schulz, M. Weigand, A. Štefančič, D. A. Mayoh, G. Balakrishnan, P. D. Hatton, and G. Schütz, Toggle-like current-induced Bloch point dynamics of 3D skyrmion strings in a room temperature nanowire, Nature Communications 13, 3630 (2022).DOI
  21. [21]A. Setescak, F. Bruckner, D. Suess, Y.-G. Choi, H. Binger, L. Boer, C. Zhang, H. Yang, C. Donnelly, U. Vool, and C. Abert, A Fourier-Space Approach to Physics-Informed Magnetization Reconstruction from Nitrogen-Vacancy Measurements, Phys. Rev. Lett. 10.1103/nc5g-3yw9 (2026), in press; preprint at arXiv:2602.17180.DOI
  22. [22]Y. Tokunaga, X. Z. Yu, J. S. White, H. M. Rønnow, D. Morikawa, Y. Taguchi, and Y. Tokura, A new class of chiral materials hosting magnetic skyrmions beyond room temperature, Nature Communications 6, 7638 (2015).DOI
  23. [23]C. Donnelly, V. Scagnoli, M. Guizar-Sicairos, M. Holler, F. Wilhelm, F. Guilloud, A. Rogalev, C. Detlefs, A. Menzel, J. Raabe, and L. J. Heyderman, High-resolution hard x-ray magnetic imaging with dichroic ptychography, Physical Review B 94, 064421 (2016).DOI
  24. [24]T. Balke, F. Davis, C. Garcia-Cardona, S. Majee, M. McCann, L. Pfister, and B. Wohlberg, Scientific Computational Imaging Code (SCICO), Journal of Open Source Software 7, 4722 (2022).DOI
  25. [25]H. Braun and A. Hauck, Tomographic reconstruction of vector fields, IEEE Transactions on Signal Processing 39, 464 (1991).DOI
  26. [26]J. L. Prince, Tomographic reconstruction of 3-D vector fields using inner product probes, IEEE Transactions on Image Processing 3, 216 (1994).pubmed.ncbi.nlm.nih.gov/18291921
  27. [27]M. Holler, M. Odstrčil, M. Guizar-Sicairos, M. Lebegue, E. Müller, S. Finizio, G. Tinti, C. David, J. Zusman, W. Unglaub, O. Bunk, J. Raabe, A. F. J. Levi, and G. Aeppli, Three-dimensional imaging of integrated circuits with macro- to nanoscale zoom, Nature Electronics 2, 464 (2019).DOI
  28. [28]C. Abert, F. Bruckner, A. Voronov, M. Lang, S. A. Pathak, S. Holt, R. Kraft, R. Allayarov, P. Flauger, S. Koraltan, T. Schref, A. Chumak, H. Fanghor, and D. Suess, NeuralMag: an open-source nodal finite-difference code for inverse micromagnetics, npj Compu- tational Materials 11, 193 (2025).
  29. [29]K. Zhu, W. Gao, Z. Wang, S. Sun, S. Huang, J. Tong, J. Li, H. Tian, Z. Li, H. Yang, Y. Zhang, O. Eriksson, F. N. Rybakov, N. S. Kiselev, and J. Li, Magnetic hopfions at room temperature (2026), arXiv:2607.26839.arxiv.org/abs/2607.26839
  30. [30]P. C. Hansen and D. P. O’Leary, The Use of the L-Curve in the Regularization of Discrete Ill-Posed Problems, SIAM Journal on Scientific Computing 14, 1487 (1993).DOI
  31. [31]D. Calvetti, S. Morigi, L. Reichel, and F. Sgallari, Tikhonov regularization and the L-curve for large discrete ill-posed problems, Journal of Computational and Applied Mathematics Numerical Analysis 2000. Vol. III: Linear Algebra, 123, 423 (2000).
  32. [32]L. D. Landau and E. M. Lifshitz, Theory of the dispersion of magnetic permeability in ferromagnetic bodies, Physikalische Zeitschrift der Sowjetunion 8, 153 (1935).
  33. [33]T. Gilbert, A phenomenological theory of damping in ferromagnetic materials, IEEE Transactions on Magnetics 40, 3443 (2004).DOI
  34. [34]K. Karube, J. S. White, V. Ukleev, C. D. Dewhurst, R. Cubitt, A. Kikkawa, Y. Tokunaga, H. M. Rønnow, Y. Tokura, and Y. Taguchi, Metastable skyrmion lattices governed by magnetic disorder and anisotropy in β-Mn-type chiral magnets, Physical Review B 102, 064408 (2020).DOI
  35. [35]F. Bruckner, S. Koraltan, C. Abert, and D. Suess, magnum.np: a PyTorch based GPU enhanced finite difference micromagnetic simulation framework for high level development and inverse design, Scientific Reports 13, 12054 (2023).arxiv.org/abs/2302.08843
  36. [36]X. Yang, H. Wang, M. Hand, K. Sawhney, B. Kaulich, I. V. Kozhevnikov, Q. Huang, and Z. Wang, Design of a multilayer-based collimated plane-grating monochromator for tender X-ray range, Journal of Synchrotron Radiation 24, 168 (2017).DOI
  37. [37]B. Enders and P. Thibault, A computational framework for ptychographic reconstructions, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 472, 20160640 (2016).DOI
  38. [38]M. Odstrčil, M. Holler, J. Raabe, and M. Guizar-Sicairos, Alignment methods for nanotomography with deep sub-pixel accuracy, Optics Express 27, 36637 (2019).DOI
  39. [39]N. Otsu, A Threshold Selection Method from Gray-Level Histograms, IEEE Transactions on Systems, Man, and Cybernetics 9, 62 (1979).DOI
  40. [40]M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-López, F. Pedregosa, and J.-P. Vert, Efficient and Modular Implicit Differentiation, in Advances in Neural Information Processing Systems, Vol. 35 (2022) pp. 5230–5242.arxiv.org/abs/2105.15183
  41. [41]G. Bézicgneul and O.-E. Ganea, Riemannian Adaptive Optimization Methods, in International Conference on Learning Representations (ICLR) (2019).arxiv.org/abs/1810.00760
  42. [42]D. P. Kingma and J. Ba, Adam: A Method for Stochastic Optimization (2017), arXiv:1412.6980 [cs].arxiv.org/abs/1412.6980
  43. [43]J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, Y. Katariya, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, and Q. Zhang, JAX: composable transformations of Python+NumPy programs (2018).

Paper details

vlnc:W_3wx3bzn8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude, Antigravity

View statement
Looking for related papers…
Contents