Quantifying Symmetry Breaking as a Design Variable for Giant Altermagnetic Spin Splitting

Problem Statement

Altermagnets represent a third class of collinear magnets characterized by zero net magnetization and momentum-dependent spin splitting (SSE) exceeding 1 eV, driven by rotational symmetries rather than inversion or translation. While magnetic point-group analysis rigorously determines whether a material can exhibit altermagnetism (a binary classification), it offers no mechanism to predict the magnitude of the spin splitting. The SSE, which governs device performance, varies continuously from ~10 meV to ~1 eV across chemically diverse materials. Current discovery workflows rely on full spin-polarized Density Functional Theory (DFT) for every candidate, creating a circular dependency where evaluation is as costly as discovery. There is a lack of a continuous, interpretable framework to map crystal structure and composition directly to SSE magnitude, limiting the field to enumerating known structure types rather than generating new design hypotheses.

Methodology

The authors propose a data-driven, inverse-design framework that replaces binary symmetry classification with continuous, physically interpretable descriptors.

1. Dataset Generation:

   • Source: 3,851 binary structures were generated using the MatterGen diffusion model, covering 323 parent compounds.
   • Labeling: Structures were labeled with DFT-computed SSE, defined as the maximum spin-channel energy separation within the occupied window $[E_F - 2 \text{ eV}, E_F]$.
   • Screening: Candidates were filtered for robust altermagnetic behavior: near-zero net magnetization ($<0.005 \mu_B$), negligible spin splitting at the $\Gamma$ point, and significant non-relativistic SSE ($\ge 1$ meV).
   • Augmentation: Each parent was expanded via isotropic and uniaxial strain sampling ($\pm 2.5\%, \pm 5\%$) to broaden the descriptor distribution without altering chemistry.

2. Descriptor Development:
   The authors introduced 52 DFT-free, motif-level descriptors derived solely from crystal coordinates and elemental identity. Three primary descriptors were identified as dominant:

   • Motif Symmetry-Breaking Index (MSBI): A continuous scalar quantifying the deviation of two antiparallel magnetic motifs from the translation ($tT$) and inversion ($PT$) relations that enforce spin degeneracy. It is calculated as the product of two dimensionless deviations:
     • $D_d$: The bond-length-normalized RMSD between ligand-vector sets of the two sublattices (matched via the Hungarian algorithm).
     • $D_I$: The deviation after point-inverting one ligand set.
     • $MSBI = D_d \times D_I$. This multiplicative form ensures MSBI is large only when both degeneracy-protecting symmetries are broken.
   • Motif Packing Fraction (MPF): The fraction of the unit cell volume (or area for planar motifs) occupied by the ligand cage, controlling the superexchange coupling scale ($J \sim t^2/U$).
   • p/d Electron Ratio: The ratio of ligand p-electrons to metal d-electrons, serving as a proxy for covalency and p–d hybridization strength within the Zaanen–Sawatzky–Allen framework.

3. Machine Learning and Optimization:

   • Surrogate Model: An XGBoost model was trained on the 3,851 DFT-labeled structures using the 52 descriptors. The model achieved a grouped cross-validated $R^2 = 0.70$ and a Mean Absolute Error (MAE) of 123.1 meV.
   • Interpretability: SHAP (SHapley Additive exPlanations) analysis was used to rank feature importance and identify the physical mechanisms driving SSE.
   • Inverse Design: Bayesian Optimization (BO) was coupled with the XGBoost surrogate to search the descriptor space for candidates maximizing predicted SSE.
   • Validation: Top candidates were mapped to real structural prototypes via cosine similarity in descriptor space and independently validated using spin-polarized DFT.

Key Results

1. Dominance of Three Descriptors:
   SHAP analysis revealed that MSBI, MPF, and the p/d electron ratio account for 37.4% of the model's total explanatory power.

   • MSBI is the dominant driver (23.2% attribution), acting as a "near-necessary gate": large SSE ($>0.4$ eV) is essentially confined to the region where MSBI $> 0.4$.
   • MPF shows a positive correlation with SSE; dense packing enhances hopping integrals and superexchange coupling.
   • p/d Ratio exhibits an inverse relationship; low ratios ($p/d < 1$) favor strong spin-asymmetric p–d hybridization and higher SSE, while $p/d \ge 1$ suppresses splitting.

2. Design Rules and Mechanisms:

   • Symmetry Breaking: MSBI quantifies the continuous degree of symmetry breaking. A case study of CuO (MSBI = 1.018, SSE = 1.634 eV) versus FeSi (MSBI = 0.0003, SSE = 0.003 eV) illustrates how the index correlates with splitting magnitude.
   • Covalency Control: A controlled comparison of VO and CrSb (sharing the same $P6_3/mmc$ host lattice) showed that changing composition to lower the p/d ratio from 1.0 to 0.6 resulted in a sevenfold SSE increase (0.165 eV to 1.194 eV), isolating the covalency mechanism.

3. Discovery of New Candidates:
   The BO workflow identified several high-SSE candidates, validated by independent DFT:

   • $\alpha$-NiS (NiAs-type): Predicted SSE 0.893 eV, DFT-validated SSE 0.823 eV. This material has been independently identified as an altermagnet, providing external cross-validation.
   • CoS (Octahedral): Predicted SSE 0.874 eV, DFT-validated SSE 1.103 eV. Independently reported as a g-wave altermagnet.
   • FeAs (Octahedral): Predicted SSE 0.867 eV, DFT-validated SSE 1.089 eV. No prior altermagnetic identification.
   • FeS (Square-planar): Predicted SSE 1.061 eV, DFT-validated SSE 1.297 eV. No prior altermagnetic identification.
   • The framework correctly identified that candidates like Cu-S and square-planar Ni-S relax to non-magnetic states, highlighting the surrogate's focus on symmetry/covalency rather than magnetic stability.

Significance and Claims

The paper claims to recast altermagnet discovery from a discrete symmetry classification problem into a continuous, interpretable structure-property design problem.

• Continuous Variable for Symmetry: The primary contribution is the MSBI, which transforms the binary question of "is altermagnetism allowed?" into a continuous variable "how strongly is it realized?" This allows the use of gradient-based and surrogate-based optimization methods.
• Actionable Design Space: By identifying MSBI, MPF, and the p/d ratio as the governing axes, the authors provide experimental handles (lattice distortion, pressure, chemical substitution) to tune SSE.
• Efficiency: The framework reduces the cost of discovery by using DFT-free descriptors to screen candidates, reserving expensive DFT calculations only for the most promising designs.
• Generalizability: The authors suggest that this descriptor logic—combining a continuous symmetry-breaking metric with geometric and covalency axes—can be extended to other symmetry-driven electronic properties beyond altermagnetism.

The work does not claim to solve thermodynamic stability or synthesis routes; rather, it asserts that if a specific coordination motif and collinear order can be stabilized, the framework predicts the resulting giant spin splitting. The validation of $\alpha$-NiS and CoS, alongside the prediction of FeAs and FeS, demonstrates the framework's ability to recover known physics and propose new motif-level design hypotheses.