Indirect prediction and screening of frustrated magnetic materials using machine learning

Technical Summary: Indirect Prediction and Screening of Frustrated Magnetic Materials Using Machine Learning

Problem Statement
The discovery of frustrated magnetic materials—systems where lattice geometry or competing exchange interactions suppress long-range magnetic ordering—is traditionally slow and resource-intensive. Magnetic frustration is quantified by the frustration index ($f = |\Theta|/T_N$), defined as the ratio of the Curie–Weiss temperature ($\Theta$) to the Néel temperature ($T_N$). Direct prediction of this index via machine learning (ML) is challenging because $\Theta$ and $T_N$ represent distinct physical phenomena (overall exchange balance vs. long-range order stabilization) and are rarely available simultaneously for the same compounds in experimental databases. Furthermore, experimental data often suffer from inconsistencies in synthesis conditions and measurement protocols, while first-principles calculations remain computationally expensive and sensitive to methodological approximations.

Methodology
The authors propose an indirect ML workflow that predicts $\Theta$ and $T_N$ independently to subsequently calculate the frustration index. The study utilizes the North East Materials Database (NEMAD), a curated dataset of experimental magnetic properties, as the training source.

1. Dataset Processing and Subdivision:

   • The initial dataset was cleaned and standardized using the Pymatgen library.
   • To address the physical heterogeneity of magnetic mechanisms, the dataset was subdivided based on chemical criteria: the presence of rare-earth (RE) elements and oxygen content. This created four classes (RE oxides, RE non-oxides, non-RE oxides, non-RE non-oxides).
   • Statistical tests (Kruskal–Wallis and Mann–Whitney) confirmed that these subclasses represent statistically distinct populations, justifying the use of separate models.

2. Descriptor Generation:

   • The study evaluated composition-based descriptors, descriptors with crystallographic symmetry, and full structural descriptors.
   • Key Finding: Composition-only descriptors provided the best predictive performance. The final feature set consisted of 146 numerical features generated via the Matminer library, including elemental property statistics (e.g., mean ground-state magnetic moment), stoichiometric descriptors, and valence orbital contributions. This approach allows screening of materials lacking resolved crystal structures.

3. Machine Learning Workflow:

   • Algorithm: The CatBoost gradient boosting algorithm was selected for its robustness with heterogeneous tabular data and ability to handle noisy descriptors.
   • Ensemble Strategy: To estimate predictive uncertainty, an ensemble of ten CatBoost models was trained using 10-fold cross-validation. The standard deviation of predictions across the ensemble served as the uncertainty metric.
   • Training: Models were optimized via randomized search with 10-fold cross-validation. Separate models were trained for $T_N$ and $\Theta$ on the full dataset and the chemically subdivided subsets.

4. Screening and Uncertainty-Aware Ranking:

   • The trained models were applied to approximately 67,000 compounds from the Materials Project database, excluding any compositions present in the training set.
   • A conservative frustration index ($f_{cons}$) was introduced to penalize uncertain predictions:
     $$f_{cons} = \frac{|\Theta| - \sigma_\Theta}{T_N + \sigma_{T_N}}$$
   • Candidates were ranked by $f_{cons}$ rather than the raw predicted index, ensuring that high-confidence candidates were selected over those with large prediction variances.

Key Results

• Model Performance:

  • Independent prediction of $T_N$ and $\Theta$ yielded moderate but physically meaningful accuracy.
  • The best-performing model for $T_N$ was the RE oxide subset ($R^2 = 0.83$).
  • The best-performing model for $\Theta$ was the RE subset ($R^2 = 0.72$).
  • SHAP analysis revealed that the mean ground-state magnetic moment was the most influential descriptor for both targets, followed by electronic structure descriptors (unfilled d- and f-shells) and crystal-chemical features.

• Screening Outcomes:

  • Screening of ~67,000 Materials Project compounds identified 20 high-confidence candidates after applying the uncertainty-aware conservative filter ($f_{cons} \geq 5$).
  • The uncertainty-aware approach significantly altered the chemical landscape of the candidates. While the raw frustration index ($f$) yielded a broad range of chemically isolated compounds, the conservative filter ($f_{cons}$) concentrated the results on chemically coherent families, specifically Sr-Ca-Ti-Mn oxide systems and ordered double perovskites.

• Validation:

  • Known Frustrated Magnets: The workflow successfully recovered known frustrated materials, including $Ba(Cr_5S_8)_2$, $Sr_2YOsO_6$, and $Sr_2ScOsO_6$. For $Ba(Cr_5S_8)_2$, the model predicted $f_{cons} = 9.68$ compared to the experimental value of $10.3$.
  • New Candidates: The remaining candidates included systems with indirect evidence of competing interactions (e.g., $Sr_2CoNbO_{6-\delta}$) and previously unexplored chemical families, such as specific Sr-Ca-Ti-Mn oxide ratios and chromium-copper selenides.

Significance and Claims
The paper claims to provide a practical framework for the accelerated discovery of frustrated magnetic materials without requiring explicit structural information or direct modeling of the frustration index.

• Indirect Prediction Strategy: By decoupling the prediction of $\Theta$ and $T_N$, the workflow respects the distinct physical origins of these parameters, leading to more interpretable models and better handling of data scarcity.
• Uncertainty-Aware Screening: The introduction of the conservative frustration index ($f_{cons}$) is highlighted as a critical step to distinguish robust candidates from unreliable extrapolations. The authors demonstrate that this approach filters out chemically isolated, high-uncertainty predictions, leaving a focused set of chemically meaningful candidates.
• Chemical Homogeneity: The study emphasizes that chemically informed dataset subdivision (specifically separating RE/non-RE and oxide/non-oxide) is essential for optimizing predictive accuracy, as magnetic mechanisms differ fundamentally across these classes.
• Scope: The authors explicitly state that the identified compounds are computationally prioritized candidates, not experimentally confirmed frustrated magnets. The workflow is designed to accelerate experimental characterization by narrowing the search space to the most promising, high-confidence targets, particularly within oxide perovskite-derived systems.