Neural Network Extracts Spin and Orbital Moments from XMCD X-ray Spectra ✛ ✦
Context
The magnetism of transition metals underpins modern technologies, from magnetic recording to spintronics. X-ray magnetic circular dichroism (XMCD) is a powerful spectroscopy tool that separately determines the spin and orbital contributions to the magnetic moment by using polarized radiation. However, the standard sum-rule analysis only operates on integral intensities and requires careful background handling, introducing systematic errors. The rapid advancement of machine learning methods opens up new possibilities for extracting hidden information from complete spectral lines.
Methods
The authors framed the task as an inverse problem: recover physical quantities from spectral shapes. To build a labeled dataset, XAS and XMCD spectra at the L₂,₃ edges of Fe, Co, and Ni were simulated using the Quanty package, which implements multiconfigurational calculations, with systematic variation of parameters: crystal field splitting (10Dq from 0.8 to 1.6 eV), spin-orbit coupling (scaling factors α₃ₔ and α₂ₚ), and exchange field (Hex up to 0.01 eV). A total of 3240 spectra were generated. A neural network with two hidden layers (256 and 128 neurons) was trained to map intensity vectors to the expectation values ⟨Sz⟩ and ⟨Lz⟩ obtained from ground-state wave functions. Numerical simulation and quantum sensing techniques were used to ensure accuracy.
Results
The trained model showed high accuracy on the test set: the coefficient of determination R² exceeded 0.99, and the root-mean-square error was below 0.02 ℏ for both moments. As revealed by the dependencies, variations in 10Dq shift spectral peaks, while Hex modulates the dichroism amplitude. The network learned to separate these contributions, effectively solving the quantum-mechanical inverse problem without relying on integral rules. Notably, the accuracy for ⟨Lz⟩ was systematically slightly higher—likely because the orbital moment has a more pronounced effect on the fine spectral structure through electromagnetic transitions.
Implications
The proposed approach does not replace sum rules but complements them by enabling information extraction from the full line profile. This is especially valuable for weak signals (e.g., in dilute systems) and in automated experiments where manual analysis becomes a bottleneck. The work bridges theoretical calculations and real measurements, paving the way for reliable determination of local magnetic moments.
Future development
The method can be extended to other 3d metals and light elements, and to experimental data accounting for instrumental broadening and noise. In the future, a universal tool for real-time XMCD spectrum analysis integrated with synchrotron radiation sources could be developed.
Impact
The work will impact condensed matter physics, spintronics, and materials science by accelerating the characterization of magnetic materials.
Next steps
Next steps include testing robustness against noise and applying to experimental spectra, as well as generalizing to non-collinear magnetic structures.
Key open problems
The research is directly connected to the long-standing problem of accurately determining spin and orbital moments from XMCD data, where traditional sum rules face limitations due to the need to account for the magnetic dipole term and background ambiguities.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors framed the analysis of XMCD spectra as an inverse problem and trained a neural network on multiconfigurational calculations to recover ⟨Sz⟩ and ⟨Lz⟩ for Fe, Co, and Ni. The model achieved R² > 0.99 and errors below 0.02 ℏ, using full XAS and XMCD spectra.
Where its strength is
The method extracts spin and orbital moments from the full line profile without relying on integral sum rules, reducing sensitivity to background and opening the door to automation.
Where it could go next
- It might be interesting to see how a similar approach transfers to recovering parameters that change the signal shape—there is a paper nearby in the archive where a neural network recovers the equation of state of neutron stars from spectral features.based on: Listening to the Stars: AI Decoder Instantly Reconstructs the Neutron Star Equation of State arXiv:2607.12722
- The idea of training a neural network to directly map spectra to physical quantities resonates with the task of rapidly determining stellar chemical compositions—there, too, the network learns on synthetic data and bypasses laborious component-by-component analysis.based on: Neural Network Reads Stars' Chemical Passport in Seconds arXiv:2604.03082
Worth trying
- Test whether adding realistic noise and backgrounds that mimic experimental conditions to the dataset improves accuracy.
- Train the network to predict not just the averages but also the variances of ⟨Sz⟩ and ⟨Lz⟩ as a measure of uncertainty.
- Explore transfer learning to spectra of other 3d metals or their alloys with minimal additional labeling.
Nearby works in our archive
- Neural Network Sculptor: Reading a Gravitational Lens from Four Pointssimilarity 0.64
- A neural network learns to see invisible polarizationsimilarity 0.61
- How a Neural Network Learned to See Quantum Foamsimilarity 0.60
- Neural Cartographer: Recovering Invisible Lawssimilarity 0.59
- Listening to the Stars: AI Decoder Instantly Reconstructs the Neutron Star Equation of Statesimilarity 0.59
- Neural Network Reads Stars' Chemical Passport in Secondssimilarity 0.58
The closest work in the archive sits at 0.64 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2607.11058
X-ray magnetic circular dichroism (XMCD) is a powerful probe of element-specific spin and orbital angular momentum. Conventional analyses based on sum rules, however, rely on integrated spectral intensities and can become insufficient when multiple parameters influence the spectral line shape. Here, we formulate XMCD analysis as an inverse problem and develop a neural-network (NN) based approach to reconstruct spin and orbital angular momentum directly from full spectral line shapes. Using many-body multiplet calculations of Fe, Co, and Ni $$L_{2,3}$$-edge X-ray absorption spectra (XAS) and XMCD spectra as a physically well-defined training dataset, we systematically vary key parameters including crystal-field splitting, spin--orbit coupling, and exchange field. The NN is trained to map spectral line shapes onto the expectation values of spin and orbital angular momentm $$\langle S_z \rangle$$ and $$\langle L_z \rangle$$, and validated using strictly test-only data. The results demonstrate accurate and unbiased reconstruction, establishing a proof of concept for data-driven inverse reconstruction from XAS and XMCD spectra. These findings show that exploiting the full XAS and XMCD line shapes provide access to information beyond conventional sum-rule analyses while remaining consistent with established theoretical frameworks.
The XMCD sum rules linking spectral integrals to magnetic moments were proposed in 1992 and are recognized as one of the cornerstones of modern magneto-optics.
Key numbers
- number of spectra: 3240
- hidden layers: 256 and 128 neurons
- 10Dq: 0.8–1.6 eV
- Hex max.: 0.01 eV
- R²: > 0.99