betaFeedback

WorksPhysics and Astronomy

Learning Magnetic Order Classification from Large-Scale Materials Databases

Abstract

The reliable identification of magnetic ground states remains a major challenge in high-throughput materials databases, where density functional theory (DFT) workflows often converge to ferromagnetic (FM) solutions. Here, we partially address this challenge by developing machine-learning classifiers trained on experimentally validated MAGNDATA magnetic materials, leveraging a limited number of simple compositional, structural, and electronic descriptors sourced from the Materials Project Database. Our propagation-vector classifiers achieve accuracies above 92%, outperforming a recent equivariant-neural-network study on a differently constructed dataset in reliably distinguishing between zero and nonzero propagation-vector structures, and exposing a systematic ferromagnetic bias inherent to the Materials Project database for more than 6840 candidate materials. In parallel, LightGBM and XGBoost models trained directly on the Materials Project labels achieve accuracies of 82% and 85%, respectively (with macro-F1 average scores of 66% and 63%), proving useful for large-scale screening for magnetic classes, when refined by MAGNDATA-trained classifiers. These results underscore the role of machine-learning techniques as corrective and exploratory tools, enabling more trustworthy databases and accelerating progress toward the identification of materials with various properties.

Introduction

Identifying the true magnetic ground state of a material remains a central challenge in condensed matter physics and materials discovery. On the computational side, ab initio methods such as density functional theory (DFT) have become indispensable for exploring magnetism, yet they face several well-documented limitations. The predicted magnetic state depends sensitively on the chosen exchange–correlation functional and on the inclusion of on-site Coulomb corrections through the Hubbard UU, which can drastically alter the relative stability of competing spin configurations [1, 2]. Furthermore, the outcome of a DFT calculation is strongly influenced by the initialization of local magnetic moments. High-throughput workflows, such as those employed in large materials databases, typically adopt a ferromagnetic starting configuration, which systematically biases the results toward FM solutions and may overlook lower-energy antiferromagnetic, ferrimagnetic, or incommensurate ground states [3]. More sophisticated ab initio techniques—including DFT+DMFT (dynamical mean-field theory) [4], hybrid functionals, and quantum Monte Carlo [5]—offer improved accuracy for correlated electron systems, but their high computational cost renders them impractical for systematic screening of large numbers of materials. Even with conventional DFT, reliably determining a material’s ground-state magnetic structure often requires exploring many candidate spin arrangements, sometimes guided by symmetry analysis or model Hamiltonians, and comparing their relative total energies [6, 7]. This procedure becomes prohibitively expensive for complex or low-symmetry systems, underscoring the difficulty of ab initio magnetic ground-state determination at scale.

Experimentally, techniques such as neutron scattering [8, 9] and, more recently, resonant inelastic X-ray scattering (RIXS) [10] and muon spin rotation (μ\muSR) [11] have provided atomic-scale insights into static and dynamic magnetic structures. Despite their unparalleled resolution, these probes depend on access to large-scale facilities — neutron reactors, spallation sources, synchrotron light sources, or muon sources — and are constrained by limited beam time availability, sample size requirements, and facility capacity. Consequently, the pace of experimental magnetic structure determination has been modest: even the most comprehensive crystallographic repositories, such as MAGNDATA, contain only about 2, 300 fully resolved magnetic structures spanning both commensurate and incommensurate orders since the 1950s [12, 13]. While continuous advances in experimental techniques (e.g., higher-flux neutron sources [14] and next-generation synchrotrons [15]) promise incremental improvements, a purely experimental mapping of magnetic materials is unlikely to keep pace with the rapidly growing demand for new magnetic compounds. This bottleneck strongly motivates the integration of complementary approaches, such as machine learning, to accelerate exploration.

Given the challenges in magnetic structure determination from both computational and experimental angles, recent research has turned toward machine learning as a powerful complementary strategy. For example, Ghosh et al. applied regression and classification models to high-quality DFT and experimental datasets of actinide compounds, achieving ∼76%\sim76\% accuracy in ordering (paramagnetic, ferromagnetic, antiferromagnetic) using a random forest classifier [16]. More recently, Merker et al. developed an equivariant Euclidean neural network that predicts both magnetic order and propagation vector directly from atomic coordinates — reporting average accuracies of ∼77.8%\sim77.8\% for magnetic order and ∼73.6%\sim73.6\% for propagation vector classification, with an impressive ∼91%\sim91\% accuracy for identifying non-magnetic structures [17]. These studies demonstrate that machine learning can effectively assist—or even partly replace—the conventional “guessing” step in DFT workflows, though fully circumventing first-principles calculations remains challenging. A complete description of magnetism remains inherently complex [18]. Here, we focus on two streamlined descriptors: magnetic order labels (FM, AFM, FiM, NM) and propagation vectors. Magnetic order labels reduce complexity into categories meaningful for applications (e.g., FM and FiM display net magnetization, AFM and NM do not), while propagation vectors capture more intricate ordering patterns when they deviate from zero. Although simplified, these descriptors offer a balanced compromise between expressiveness and tractability—serving our goal of scalable, interpretable predictions.

In this paper, our goal is to build ML classifiers that can predict the magnetic order (FM, FiM, AFM, NM) on the Materials Project (MP) database [3, 19], a binary information about the propagation vector (zero or nonzero) on Bilbao Crystallographic Server (BCS)’s MAGNDATA [12, 13], and finally try to expose the FM bias on the MP database through the efficient MAGNDATA-trained classifier. Although these outputs do not cover a full description of the magnetic order, they still contain critical information on the existence of local magnetic moments, the magnitude of the net magnetization, and the relative spin alignments, all of which are highly informative for distinguishing between different magnetic classes.

Data Collection & Pre-processing

Our strategy is to use a few simple descriptors that are linked either to the formation of magnetic moments or the interactions between them, which eventually lead to the formation of the long-range magnetic order. Therefore, we consider three classes of descriptors: compositional, structural, and electronic. The compositional part contains the different constituent chemical elements that form the compound, and therefore implicitly encoding information about their likely oxidation states and electronic configurations, which in turn affect the likelihood of forming local magnetic moments. The chemical composition was implemented using a one-hot encoding scheme, where each element is represented by a unique basis vector in a high-dimensional binary space. The structural part consists of the crystal system (cubic, tetragonal, orthorhombic, hexagonal, trigonal, monoclinic, and triclinic), the atomic density, volume of the unit cell (which, combined with the atomic density, relates to the number density), and mass density, which is indirectly related to the atomic masses of the constituent elements. These features are linked to the symmetry of the crystal, the packing and compactness of the structure, and the lattice geometry, which in turn influence the superexchange and direct exchange pathways that govern magnetic interactions.

Electronic descriptors — such as the band gap, conduction band minimum (CBM), valence band maximum (VBM), and Fermi energy — supplement this structural information by characterizing the degree of electron localization and the availability of itinerant carriers. In particular, materials with large band gaps and partially filled d- or f-shells typically exhibit localized electrons, which favor the formation of robust local moments and superexchange-driven antiferromagnetic or ferrimagnetic order [20, 21, 22]. In contrast, metallic systems often support itinerant magnetism [23, 24] or spin-density-wave states [25, 26]. Exceptions exist, including correlated systems where local and itinerant behavior coexist [27, 28]. By linking electron localization and carrier availability to spin interactions, these descriptors are essential predictors of the type and stability of magnetic order in crystalline materials. It is worth noting that some of the descriptors we adopt do not give useful information unless combined with other features such as VBM, CBM, Fermi energy or volume. In keeping with the goal of demonstrating that even a coarse-grained descriptor set is sufficient to train our classifiers, we deliberately restrict our feature set to the descriptors discussed above. Other meaningful descriptors (e.g. electronegativity and atomic dipole polarizability that are considered in Ref. [17]) are intentionally excluded to keep the pipeline as simple as possible and strictly defined by what an end-user can easily pull from the MP database itself without further pre-processing. Such descriptors could be related indirectly to the compositional and electronic features that we employ. To make this quantitative, we will later relax this restriction and quantify the effect of including such descriptors on classifier performance.

For the MP classifier, we construct a dataset of the descriptors and the magnetic class for each material through the Materials Project API. This results in 154,803 materials, with 104,795 unique ones as there are materials with more than one MP profile. Due to continuous updates of the MP, the number may change upon a new access. We focus on materials with at least one magnetic element such as transition metal elements (Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Y, Nb, Mo, Ru, Rh, Re, Os, Ir, and Pt), lanthanides (Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Ho, Er, Tm, and Yb) and actinides (Th, U, Np, and Pu). The magnetic order classification on the MP is determined from spin-polarized DFT calculations and post-processed with the CollinearMagneticStructureAnalyzer in the pymatgen library. A magnetic class is assigned based on the total magnetization of the unit cell and the alignment of different spins. If all spins align with nonzero net magnetization, the compound is labelled FM, if spins anti-align with nonzero net magnetization, it is labelled FiM, and if total magnetization vanishes but the local moments survive, it is AFM. If both the total magnetization and the local moments vanish, then it it NM. Based on this scheme, a magnetic class is assigned to each material on the MP.

The MP ordering is not derived from experiment but from the sign of the total magnetic moment in a single spin-polarized DFT calculation, initialized with a ferromagnetic guess (all MAGMOM values set nonzero and parallel). For compounds containing no magnetic ions, the self-consistent moment should relax to zero; in practice, however, a small residual moment often persists due to finite electronic smearing, incomplete k-point convergence, and the absence of an explicit antiferromagnetic reference calculation for comparison. When the magnitude of the total moment exceeds MP’s internal threshold, the compound is classified as FM even when no physical mechanism for ferromagnetism exists. This is a documented limitation of MP’s magnetic classification scheme — see [29], which describes the MP magnetic workflow in detail and reports that the concordance between MP magnetic ordering labels and experimental MAGNDATA classifications is limited, with compounds of weak or numerically-induced moments constituting a known class of mislabels. Finally, the magnetic class labels depend mainly on the used DFT functional (GGA or GGA+UU) which can be sensitive the value of the Hubbard UU. As a result of these limitations, the MP dataset shows high bias towards the FM and NM classes while weakly capturing FiM and AFM classes, as shown in the histogram distribution in Figure 1. Before adding the filter of the existence of magnetic elements, the NM class dominates with the FM class right behind it, while the filter produces an opposite distribution, as shown in Figure 1 (b). The different magnetic classes of the filtered set of materials shows some distinct distributions over the selected descriptors as presented in Figure 2 where we select two electronic features: the band gap and the valence band maximum, and two structural features: the density and the atomic density. Therefore, while useful for a large-scale screening, the magnetic class labels on the MP should be treated as heuristic rather than conclusive regarding describing the true magnetic ground state.

Image not extracted

Figure 1. (a) Distribution of magnetic classes in the Materials Project database without filtering. (b) Distribution of magnetic classes restricted to Materials Project compounds containing at least one magnetic element.

Image not extracted

Figure 2. Distributions (log-scale counts) of selected features in the refined Materials Project dataset, restricted to compounds containing at least one magnetic element and resolved by magnetic class: (a) electronic band gap, (b) valence band maximum (VBM), (c) mass density, and (d) atomic density. These descriptors highlight systematic differences between ferromagnetic (FM), antiferromagnetic (AFM), ferrimagnetic (FiM), and non-magnetic (NM) compounds.

Our second database, MAGNDATA on the BCS, is

the most comprehensive experimentally-derived magnetic structures database through neutron diffraction studies. Although these experiments provide reliable description of magnetic orders, they are difficult to perform over a large scale and therefore it contains only 2,167 commensurate magnetic materials (at the time of the data collection). On the contrary of the MP dataset, the distribution of the magnetic classes in MAGNDATA shows the dominance of the AFM with 1,207 collinear AFM and 851 non-collinear AFM, as shown in Figure 3 (a). This extreme imbalance makes it difficult to train a ML model given the rare data points about the FM and FiM classes. Therefore, since the distribution of the magnetic propagation vector (zero vs. nonzero) is balanced, as shown in Figure 3 (b), we choose it as a useful piece of information about the magnetic structure to train our ML models over. Note that we merged materials of more than one propagation vector into the nonzero class for simplicity to train a binary classifier over a balanced dataset. Adopting the same set of descriptors, we supplement the MAGNDATA dataset with the corresponding features from the MP database. This results in 3,980 entries, with redundancy stemming from the existence of more than one MP-profile for a number of materials.

Image not extracted

Figure 3. Distributions of magnetic classes in the MAGNDATA dataset: (a) magnetic order labels, including ferromagnetic (FM), ferrimagnetic (FiM), antiferromagnetic (AFM), and complex orders; (b) propagation vector labels, distinguishing between zero and nonzero kk-vectors. These distributions highlight the prevalence of AFM and complex orders in experimentally determined magnetic structures, as well as the near balance between commensurate and incommensurate propagation vectors.

Data flow and leakage controls. In the rest of the paper, we use two distinct classifier families, each with its own scope. The MP classifiers are trained and evaluated on the MP database (more details in the next section), and are only applied to the MP database. Since training, evaluation, and application all come from the same database, there is no cross-database leakage to control for, and MAGNDATA entries are not removed from the MP pipeline. The MAGNDATA classifiers are trained and evaluated on MAGNDATA itself; the reported performance metrics measure generalization within MAGNDATA. We then apply these MAGNDATA-trained classifiers to MP to flag entries whose FM label are inconsistent with the learned propagation-vector predictions. This cross-database step is where leakage would matter: some MP entries are also in MAGNDATA, so applying the classifiers to them would amount to re-predicting materials the models had already seen during training. To prevent this, we remove all MAGNDATA materials from MP before the cross-database application, so the flagged list contains only MP entries that were never part of the MAGNDATA training data.

Machine Learning Classifiers

Now, we train ML classifiers for each of our datasets independently after partitioning each randomly with a stratified strategy (to take into account the imbalance in the MP dataset) into training, validation, and testing subsets with ratios of 70%, 20%, 10%, respectively. Within the training set, we employed a multi-fold stratified cross validation (CV) to tune hyperparameters and to obtain a robust estimate of the model stability across the different folds. The separate validation set is used to compare the performance of the different trained models while the final test set is a final check set that is used only once for further confirmation of no overfitting (beside the model stability across the different folds in the CV employed during the training). It is worth noting that our ML classifier on the MAGNDATA propagation vector has two main goals. First is to get an accurate description of the propagation vector itself which reveals some information about the underlying magnetic structure. The second is to apply our most efficient propagation vector classifier to the MP dataset. This enables us to partially address the FM bias on the MP database through predicting a nonzero propagation vector, which excludes the possibility of a FM order.

Performance Metrics. Throughout this work we evaluate the ML classifiers using three complementary metrics: accuracy, precision, and the F1F_1 score [30, 31]. For a given class, let TP (true positives) denote the number of materials correctly assigned to that class, FP (false positives) the number of materials incorrectly assigned to it, FN (false negatives) the number of materials that belong to the class but were assigned elsewhere, and TN (true negatives) the number of materials correctly assigned to another class. The first metric is accuracy which is the overall fraction of correct predictions,

Accuracy=TP+TNTP+TN+FP+FN,(1)\mathrm{Accuracy} = \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}, \tag*{(1)}

and while intuitive, it is misleading on imbalanced datasets: a classifier that labels every material as FM would achieve high accuracy on an FM-heavy training set despite learning nothing useful. To avoid this, we rely more heavily on precision and recall, defined for each class as

Precision=TPTP+FP,Recall=TPTP+FN.(2)\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad\mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}. \tag*{(2)}

Precision measures how often the classifier is correct when it flags a material as belonging to the class, while recall measures what fraction of the materials that actually belong to the class the classifier successfully identifies. The F1F_1 score combines these two through their harmonic mean,

F1=2⋅Precision⋅RecallPrecision+Recall,(3)F_1 = 2 \cdot\frac{\mathrm{Precision}\cdot\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}, \tag*{(3)}

and takes values between 0 (worst) and 1 (perfect classification). As a rough guide, a classifier that is not learning anything useful—for example, one that always predicts a single class—typically produces F1≲1/kF_1 \lesssim1/k for a balanced kk-class problem (e.g. around 0.25 for our four-class MP task and 0.5 for the binary MAGNDATA task), while a classifier that successfully separates the classes produces F1F_1 values substantially above this baseline.

Because our MAGNDATA and MP datasets are class-imbalanced, we report the macro-averaged F1F_1 score, defined as

F1macro=1k∑i=1kF1(i),(4)F_1^{\mathrm{macro}} = \frac{1}{k}\sum_{i=1}^{k} F_1^{(i)}, \tag*{(4)}

where F1(i)F_1^{(i)} is the F1F_1 score of class ii and kk is the number of classes. This gives each class equal weight regardless of its population, and therefore ensures that minority classes (AFM, FiM) contribute to the reported score as much as the majority classes (FM, NM), preventing a classifier that performs well only on majority classes from appearing deceptively strong overall. The F1F_1 score and its macro-averaged variant are standard metrics in the machine-learning literature [30, 31, 32] and are widely used for imbalanced classification tasks in materials science applications. What qualifies as a high F1F_1 is task-dependent and varies with the inherent difficulty of the classification problem; for the magnetic-classification tasks considered in this work, a non-informative classifier (e.g. one that predicts labels at random in proportion to their class frequencies) produces F1macro=1/kF_1^{\mathrm{macro}} = 1/k regardless of class imbalance (around 0.25 for our four-class MP task, see e.g. the Dummy Classifier in Table I, and 0.5 for the binary MAGNDATA task).

ML Magnetic Ordering Class on The Materials Project

As a first task, we benchmarked several supervised ML models, using scikit-learn module [32], on the MP dataset: k-Nearest Neighbors, Decision Tree, LightGBM, Support Vector Machine, Random Forest, and XGBoost. To establish a true lower bound, we used scikit-learn’s DummyClassifier with stratified strategy, which predicts labels at random, weighted by the class proportions in the training set. This baseline is intentionally feature-agnostic and its performance reflects only the dataset’s class imbalance

ClassPrecisionRecallF1-score
AFM0.050.050.05
FM0.510.520.51
FiM0.100.100.10
NM0.360.360.36
Accuracy0.40
Macro avg0.260.260.26
Weighted avg0.400.400.40

Table I. Baseline performance of the stratified DummyClassifier on the Materials Project validation set.

independently from the features, and it is evaluated under the same 5-fold stratified CV, 20% validation, and 10% test protocol as our real models (same splits, fixed random_state). The overall accuracy equals ∑kpk2\sum_k p_k^2 where pkp_k is the kk–th class prior, and its per-class recall and precision are equal to pkp_k. This results in a per-class F1k=pkF_{1k} = p_k and a macro-F1F_1 of 1/C1/C where CC is the number of classes. Consequently, any model that does not outperform the DummyClassifier’s accuracy (∑kpk2\sum_k p_k^2) and macro F1F_1 of 1/C1/C captures no information beyond the classes priors. We check each of our ML models performance across the different CV folds to ensure that our results are reproducible and not artifacts of a particular data partition. For the MP dataset, performance of the baseline model is summarized in Table I.

As shown in Figure 4, all trained classifiers substantially outperformed the chance-level baseline model (accuracy = 0.40, macro F1F_1 = 0.26) demonstrating that the features capture meaningful information about the underlying magnetic order. Simpler models such as a single decision tree (accuracy = 0.75, macro F1F_1 = 0.57) and k-nearest neighbors (accuracy = 0.80, macro F1F_1 = 0.61) improved considerably over the baseline but remained below the ensemble-based methods. Support vector machines achieved only a moderate performance (accuracy = 0.76, macro F1F_1 = 0.55), likely due to the nonlinearity and the imbalance of our data. However, ensemble tree methods achieved the strongest results: XGBoost and Random Forest both reached accuracies above 84% with macro F1F_1 average scores around 0.63 and 0.62, respectively, while LightGBM delivered the highest macro F1F_1 score of 0.66 at 0.82 accuracy. Hyperparameter optimization for all classifiers was performed using RandomizedSearchCV or GridSearchCV with the search space and the selected hyperparameters reported in Table VII. Given the balance between the overall correctness (captured by accuracy) and fairness across classes highlights LightGBM and XGBoost as the top performing models, with LightGBM showing superior performance on under-represented classes, and XGBoost providing the best aggregate predictive power. The strong performance demonstrated here using gradient-boosted tree ensembles in our study shows consistency with recent literature. For example, He et al. [33] demonstrated the effectiveness of XGBoost in predicting the correct topological class of a given material, after being trained on the topological materials databases, with an accuracy of 85.2%. Additionally, Ghosh et al. [16] reported a Random Forest classifier for magnetic order of uranium compounds with a mean accuracy of 60.2%. These results collectively reinforces the powerfulness of the tree-based ML methods on learning complex material properties such as magnetism and topology using purely structural or derived features.

Image not extracted

Figure 4. Performance of different classifiers on the Materials Project validation set, measured by accuracy (blue) and macro-F1F_1 score (green). Simple baselines such as the dummy and decision tree classifiers perform significantly worse than ensemble methods, with Random Forest, LightGBM, and XGBoost achieving the highest accuracies (0.82–0.85) and balanced F1F_1 scores (0.62–0.66). Results shown here correspond to the four-class task for materials with at least one magnetic element, as discussed in the main text.

HyperparameterSearch spaceSelected (best)Notes (meaning)
XGBoost — RandomizedSearchCV; mean CV F1macro=0.623F_1^{\mathrm{macro}} = 0.623
n_estimatorsrandint(250, 750)647number of boosting rounds
max_depthrandint(3, 11)9maximum tree depth
learning_rateloguniform(0.03, 0.25)0.139shrinkage per round
subsample{0.7, 0.85, 1.0}1.0row subsampling fraction
colsample_bytree{0.6, 0.8, 1.0}1.0column subsampling per tree
min_child_weightrandint(1, 8)7min. sum of instance weight in child
gammarandint(0, 6)0minimum loss reduction for split
reg_lambdaloguniform(10−110^{-1}, 10110^{1})1.110L2 regularization
reg_alphaloguniform(10−410^{-4}, 10−110^{-1})0.003L1 regularization
LightGBM — RandomizedSearchCV; mean CV F1macro=0.658F_1^{\mathrm{macro}} = 0.658
num_leavesrandint(31, 129)120maximum leaves per tree
max_depth{6, 8, 10, 12}8maximum tree depth
min_child_samplesrandint(40, 201)43min. samples per leaf
min_split_gainloguniform(10−210^{-2}, 1.0)0.023minimum gain to split
subsample{0.7, 0.8, 0.9}0.8row subsampling fraction
bagging_freq{1}1bagging frequency (forced on)
colsample_bytree{0.6, 0.8}0.8column subsampling per tree
learning_rateloguniform(0.03, 0.15)0.069shrinkage per round
n_estimatorsrandint(250, 800)651number of boosting rounds
reg_lambdaloguniform(1.0, 20.0)3.267L2 regularization
reg_alphaloguniform(10−210^{-2}, 1.0)0.079L1 regularization
Random Forest — GridSearchCV; mean CV F1macro=0.614F_1^{\mathrm{macro}} = 0.614
n_estimators{100, 250, 300, 350, 400}300number of trees
max_depth{15, 25, 35, 50}35maximum tree depth
min_samples_split{2, 5}5min. samples to split a node
class_weight{balanced, None}balancedclass-imbalance handling
Decision Tree — GridSearchCV; mean CV F1macro=0.561F_1^{\mathrm{macro}} = 0.561
max_depth{14, 16, 18, 22, 30, 36, 40}30maximum tree depth
min_samples_leaf{1, 3, 5, 8, 10}1min. samples per leaf
class_weight{balanced}balancedclass-imbalance handling
Linear SVM — RandomizedSearchCV; mean CV F1macro=0.536F_1^{\mathrm{macro}} = 0.536
Cloguniform(10−310^{-3}, 10210^{2})0.985inverse regularization strength
loss{squared_hinge, hinge}squared_hingeloss function
class_weightbalanced (fixed)balancedclass-imbalance handling
max_iter5000 (fixed)5000solver iteration cap
kNN — RandomizedSearchCV; mean CV F1macro=0.596F_1^{\mathrm{macro}} = 0.596
n_neighborsrandint(3, 26)4number of neighbors
weights{uniform, distance}distanceneighbor weighting
metric{euclidean, manhattan, cosine}manhattandistance metric
algorithm{brute}brutenearest-neighbor algorithm

Table VII. Hyperparameter tuning summary for the MP classifiers on the four-class case for materials with at least one magnetic element. For each classifier we list the hyperparameters tuned, the search space explored, the search method and number of iterations, and the final selected hyperparameter values (best). All searches used 5-fold stratified cross-validation with f1 macro as the scoring metric. “randint(aa, bb)” denotes a uniform integer distribution on [a,b)[a,b); “loguniform(aa, bb)” denotes a log-uniform continuous distribution on [a,b][a,b].

As shown in the confusion matrix in Figure 5 (a), the LightGBM classifier recognizes the FM and NM classes with a high fidelity (F1F_1 = 0.84 and 0.93, respectively), while the AFM and FiM classes remain more challenging (F1≃0.45F_1 \simeq0.45). This gap primarily arises from confusion between FM and FiM classes and the partial overlap between AFM and FM likely stemming from the fact that many materials in these classes share similar structural and compositional features, making them harder to distinguish using our current set of simple descriptors. In contrast, NM materials are readily identified with an F1F_1 score of 0.94 showing that the model sees distinct features for this class.

Image not extracted

Figure 5. Confusion matrices for the LightGBM classifier applied to the Materials Project dataset (restricted to compounds containing at least one magnetic element), shown under two labeling schemes: (a) four-class classification (FM, AFM, FiM, NM) and (b) binary classification (magnetic vs. nonmagnetic). Panel (a) corresponds to the classification performance reported in Figure 4.

Now we compare the performance of LightGBM under different class groupings. When FM and FiM are merged into a single class FM_FiM, the confusion matrix in Figure 6 (a) shows an improvement: the majority of the FM_FiM and NM materials are recovered with an accuracy over 90%, although AFM remains challenging to identify. Finally, in the binary setting of magnetic vs. nonmagnetic, shown in Figure 5 (b), LightGBM achieves a remarkable separation with over 95% correct classification of both classes. Again, this trend reflects the physical similarities of FM and FiM on one hand, and the scarcity of the AFM compounds on the MP database.

Image not extracted

Figure 6. Confusion matrices for: (a) the LightGBM classifier independently re-trained under a three-class labeling scheme (AFM, FM/FiM, NM) on the same Materials Project dataset, and (b) the result of Merker et al. [17]. The three-class scheme in panel (a) is adopted specifically for direct comparison with [17]; small numerical differences relative to Figure 5(a), which uses the original four-class scheme, and to Figure 5(b), which uses the two-class scheme, reflect the independent model learning (e.g. different stratified splits and independently-tuned hyperparameters) under the two labeling schemes.

To gauge our LightGBM relative to recent literature, we compare our 3-class model (FM_FiM, AFM, NM) with the results of Merker et al. [17], as shown in Figure 6. It is critical to note that the two studies are working on different datasets: Merker et al. [17] considers only compounds with calculations restricted to generalized gradient approximation and Hubbard interaction (GGA + U) [3, 19], resulting in a small dataset. In both, the recovery of the AFM class remains challenging, frequently interfering with the FM_FiM class. Again, this reflects the fundamental challenge of distinguishing the AFM from the FM and FiM orders based solely on global descriptors. However, our LightGBM performs noticeably higher in both the FM_FiM and NM classes, with classification rates of 90.8% and 93.1%, respectively, compared to recovery rates of 69.80% and 93.95%, respectively, while Merker et al. [17]’s model shows better performance on the AFM class with a recovery rate of 67.47% compared to ours of 48.1%.

Feature importance. The feature importance analysis was performed using LightGBM’s gain-based importance measure, which sums the reduction in the loss function contributed by each given feature across all decision-tree splits. The importance of a feature ff is defined as:

Igain(f)=∑s∈S(f)ΔL(s)I_{\mathrm{gain}}(f) = \sum_{s \in S(f)} \Delta L(s)

where S(f)S(f) is the set of all splits that employ the feature ff, and ΔL(s)\Delta L(s) is the improvement achieved for the loss function through this split. Therefore, this measure provides a distinction for features that have the strongest predictive power rather than those that are not used frequently. The results show that elemental composition dominates, with around one third of the total model gain (see Figure 7 (a)). Among the numerical features, the electronic band gap shows high importance, possibly due to its central role in determining the electronic ground states and their connection to magnetic order. Structural-related descriptors such as the atomic and overall density and volume, along with other electronic features such as CBM, VBM, and Fermi energy, also rank highly. This underscores the fundamental interplay between structural packing, electronic filling, and magnetic order. Crystal system contributes only a few percent, suggesting that global crystal type does not efficiently distinguish magnetic orders alone without extra features. This originates from the fact that the crystal structure (cubic, tetragonal, orthorhombic, etc.) is a coarse-grained symmetry classification that captures the metric of the unit cell and constrains the allowed point-group symmetries of the unit cell but discards information about the local coordination of magnetic ions and the bond-angle distributions that govern superexchange (per the Goodenough-Kanamori-Anderson rules [21]). Two compounds in the same crystal system can therefore exhibit very different magnetic ground states, and compounds in different crystal systems can share a common magnetic mechanism. Individually examined, the elemental contribution is dominated by 3d transition metals such as Mn, Fe, Co, Cr, and Ni, with oxygen also highly performing (see Figure 7 (b)), possibly due to its role in mediating superexchange pathways in many compounds in the database. This physically reflects the fact that transition-metal chemistry and their surrounding anion environment being critical to the magnetic interactions and therefore the emergent magnetic order.

Image not extracted

Figure 7. Feature importance analysis based on LightGBM’s gain metric, which aggregates the reduction in the loss function contributed by each feature across all decision-tree splits. (a) Grouped feature importances, highlighting the dominant role of compositional descriptors (elements), electronic band gap, and atomic density. (b) Top 15 elemental contributions, showing Mn, Fe, and O as the most influential elements for classification performance.

However, it is important to distinguish between two qualitatively different ways in which a feature can be informative to a tree-ensemble classifier. A feature may be predictive because it directly encodes a physical quantity that influences the target — for example, the band gap, Fermi level, unit-cell volume, or site count, each of which reflects the underlying electronic structure, hybridization strength, and density of magnetic centers relevant to magnetic ordering. Alternatively, a feature may be predictive only as a statistical proxy for a chemical or structural family that is over- or under-represented in the training data, with no direct physical mechanism linking it to the target. Feature-importance rankings from tree ensembles conflate these two modes and cannot, by themselves, distinguish them. We therefore caution the reader against inferring a physical role from such rankings without an independent mechanism-based justification which is beyond the scope of this work. Evaluating the performance of LightGBM over the final held-out test set (10%), shown in Table II, results in a classification accuracy of 83% and a macro F1F_1 average score of 69%, which closely matches the performance over the validation set of 82% accuracy and 66% macro F1F_1 average score. The consistency of the classifier’s performance over the two subsets indicates that the model generalizes well with no evidence of overfitting to the training distribution. Class-level metrics are also preserved: NM and FM classes are highly recovered, while AFM and FiM classes achieve lower F1F_1 scores, but their recall and precision remain comparable across the two subsets. Stability of the performance is confirmed as well during the five-fold cross validation strategy adopted at the training stage. This stability supports the robustness of the LightGBM classifier and its applicability to unseen compounds.

ClassValidation Prec.Validation Rec.Validation F1Test Prec.Test Rec.Test F1
AFM0.420.480.450.460.540.50
FM0.850.820.840.850.830.84
FiM0.410.470.440.450.500.47
NM0.940.930.930.930.930.93
Overall acc.0.820.83
Macro avg0.650.670.660.680.700.69
Weighted avg0.830.820.820.830.830.83

Table II. LightGBM performance by class on the validation and test sets. Values reported to two decimals.

ML Propagation Vector on MAGNDATA

For the second classification task, we focus on the second dataset formed by supplementing the materials extracted from MAGNDATA with features from the MP database, as described earlier. We train two classifiers: Random Forest and XGBoost on learning the binary propagation vector (zero vs. nonzero). Even though this binary information is simpler than distinguishing different magnetic orders, it is still physically meaningful since a nonzero propagation vector implies a modulated or complex magnetic structure, which shall also help us in the next section to diagnose the FM bias on the MP database. We adopt the same data splitting strategy as discussed earlier with the same five-fold cross validation strategy at the training step. Both models achieved strong performance on the validation set, with XGBoost reaching an overall accuracy of 92% and a macro F1F_1 average score of 90%, while Random Forest slightly outperforms it with an accuracy of 93% and 91% macro F1F_1 average score. Hyperparameter optimization is reported in Table VIII. In both models, the zero propagation vector class was classified with a very high recall and precision (around 94%), while the nonzero case showed slightly lower but still robust with a recall of 88–90% and a precision of 81–82%. As shown in the confusion matrix in Figure 8, the majority of the error stems from mislabeling a small fraction of zero-kk cases as nonzero-kk materials, while the nonzero-kk structures are recovered with higher reliability. Highlighting Random Forest as the best propagation vector classifier (even though it slightly outperforms XGBoost), we utilize its feature importance analysis (see Figure 10) to show that the chemical composition of a given compound provides the largest contribution, followed by structural descriptors such as density and atomic density, and electronic descriptors such as band gap, CBM, and VBM. The crystal system still shows lower predictive power. Finally, the model generalizes well to unseen materials, as confirmed in the five-fold cross-validation process, and its performance over the final test set. To benchmark the strong performance of our Random Forest, we compare it to the results of [17], as shown in Figure 9. A similar analysis holds for the XGBoost classifier.

Image not extracted

(a)

Image not extracted

(b)

Figure 8. Confusion matrices for propagation-vector classifiers on MAGNDATA: (a) Random Forest and (b) XGBoost.

Image not extracted

Figure 9. Comparison of propagation vector classification results (zero vs. nonzero) shown as row-normalized confusion matrices with percentage values for (a) our Random Forest model trained on MAGNDATA and (b) results from Merker et al. [17].

Image not extracted

Figure 10. Random Forest feature importance analysis for propagation vector classification on the MAGNDATA dataset. Grouped feature importances, showing that compositional descriptors (elements) dominate, followed by density-related and electronic features such as atomic density, band gap, and conduction/valence band edges.

HyperparameterSearch spaceSelected (best)Notes
Random Forest — RandomizedSearchCV; mean CV F1macro=0.883F_1^{\mathrm{macro}} = 0.883
n_estimatorsrandint(150, 801)405number of trees
max_depth{8, 12, 16, 20}16maximum tree depth
min_samples_splitrandint(2, 21)2min. samples to split a node
min_samples_leafrandint(1, 11)1min. samples per leaf
max_features{sqrt, log2, None}Nonefeatures considered per split
bootstrap{True, False}Truewhether to bootstrap samples
XGBoost — RandomizedSearchCV; best CV F1macro=0.879F_1^{\mathrm{macro}} = 0.879
n_estimatorsrandint(150, 801)428number of boosting rounds
max_depthrandint(3, 17)15maximum tree depth
learning_rateloguniform(0.01, 0.30)0.215shrinkage per round
subsample{0.7, 0.8, 0.9, 1.0}1.0row subsampling fraction
colsample_bytree{0.6, 0.8, 1.0}0.8column subsampling per tree
min_child_weightrandint(1, 8)1min. sum of instance weight in child
gammaloguniform(10−310^{-3}, 1.0)0.021min. loss reduction for split
reg_lambdaloguniform(10−210^{-2}, 10.0)2.841L2 regularization
reg_alphaloguniform(10−310^{-3}, 1.0)0.020L1 regularization

Table VIII. Hyperparameter tuning summary for the MAGNDATA-trained propagation-vector classifiers (binary task: zero-kk vs. nonzero-kk). For each classifier we list the hyperparameters tuned, the search space explored, the search method and number of iterations, and the final selected hyperparameter values. All searches used 6-fold stratified cross-validation with f1f1 macro as the scoring metric. “randint(a,ba,b)” denotes a uniform integer distribution on [a,b)[a,b); “loguniform(a,ba,b)” denotes a log-uniform continuous distribution on [a,b][a,b].

MAGNDATA ML Classifiers Applied onto the Materials Project Database

In our final task, we employ our propagation vector classifiers trained on MAGNDATA to the MP database in order to assign each compound a binary propagation vector label. This enables a large-scale prediction of the binary kk-vector, a goal that is not achievable on the MP database alone. Through the predictions of this classifier, our goal is to identify compounds that are labeled as ferromagnets on the Materials Project database, yet predicted to have a nonzero kk-vector through the MAGNDATA-trained-on classifiers. Before that, we filter out MAGNDATA materials that are available on the MP database to avoid data leakage as discussed in section II. We also use the MP dataset without filtering out materials that do not have at least one magnetic ion in order to test the MAGNDATA classifiers capability of capturing such cases. Applying the kk-vector Random Forest classifier, we identify 16, 973 materials labeled as ferromagnets on the Materials Project database predicted to have a non-zero propagation vector. Due to the existence of more than one MP profile for some materials on the MP database, among these are 10, 354 unique materials (see Supplementary Table 2). Repeating this for the kk-vector XGBoost, 14, 064 (among them are 9, 383 unique compounds, see Supplementary Table 3) MP-labeled FMs are predicted to have a nonzero propagation vector. Given the strong performance of the kk-vector classifiers on MAGNDATA, which is a realistic database compared to the DFT-based Materials Project database, we rely on the mentioned contradiction above − having a predicted nonzero kk-vector and a FM label on the MP database − to predict that these materials are candidate wrongly-labeled FMs on the MP database, therefore partially address its bias towards the FM class. As a practice of having more confidence in the final set of potentially mislabeled MP-ferromagnets, we choose the intersecting set of compounds of 7, 843 (see Supplementary Table 1) between the two kk-vector classifiers. A few examples of our tabulated candidate MP-mislabeled FM materials include FeCl₃ [40], VBr₃ [53], VF₃ [54], α-Gd₂S₃ [41], MnSe [48], Mn₅O₈ [47], and Mn₂O₃ [46] (see Table III for more materials), where the associated references validate our predictions of the existence of non-ferromagnetic orders.

MaterialReferenceComments
Ba2GdNbO6\mathrm{Ba}_{2}\mathrm{GdNbO}_{6}Ref. [34]Predominantly paramagnetic down to 1.8 K.
Ba2GdSbO6\mathrm{Ba}_{2}\mathrm{GdSbO}_{6}Ref. [35]Heat capacity measurements indicate a lack of magnetic ordering down to 0.4 K.
Ba2GdTaO6\mathrm{Ba}_{2}\mathrm{GdTaO}_{6}Ref. [36]Magnetic susceptibility measurements shows paramagnetism down to 5 K.
CaV4O9\mathrm{CaV}_{4}\mathrm{O}_{9}Ref. [37]Two-dimensional spin-gap antiferromagnet.
Co2PO4F\mathrm{Co}_{2}\mathrm{PO}_{4}\mathrm{F}Ref. [38]Antiferromagnetic with TN≈14T_N \approx14 K and negative Curie–Weiss temperature.
CsCr(MoO4)2\mathrm{CsCr(MoO}_{4}\mathrm{)}_{2}Ref. [39]Springer Materials classifies it as AFM.
FeCl3\mathrm{FeCl}_{3}Ref. [40]Single-crystal neutron diffraction and SANS on a non-centrosymmetric polymorph reveal an antiferromagnetic transition at TN=8.6T_N = 8.6 K with propagation vector k=(1/2,0,1/3)\mathbf{k} = (1/2, 0, 1/3).
α\alpha-Gd2S3\mathrm{Gd}_{2}\mathrm{S}_{3}Ref. [41]Orthorhombic (PnmaPnma) Gd3+\mathrm{Gd}^{3+} sulfide; magnetic susceptibility of single-crystal samples shows long-range antiferromagnetic ordering below ∼10\sim10 K.
HCrO2\mathrm{HCrO}_{2}Ref. [42]Triangular-lattice antiferromagnet
KCrS2\mathrm{KCrS}_{2}Ref. [43]Magnetic structure consists of ferromagnetic sheets coupled antiferromagnetically.
KFe(MoO4)2\mathrm{KFe(MoO}_{4}\mathrm{)}_{2}Ref. [44]Stacked triangular antiferromagnet; TN≈2.5T_N \approx2.5 K; coexistence of one collinear AFM subsystem and one spiral subsystem.
LaCrTeO6\mathrm{LaCrTeO}_{6}Ref. [45]Springer Materials classifies LaCrTeO6_6 as AFM.
Mn2O3\mathrm{Mn}_{2}\mathrm{O}_{3}Ref. [46]Mn2O3\mathrm{Mn}_{2}\mathrm{O}_{3} exhibits two intricate antiferromagnetic structures below ∼100\sim100 K (neutron powder diffraction).
Mn5O8\mathrm{Mn}_{5}\mathrm{O}_{8}Ref. [47]Magnetic susceptibility and ESR show antiferromagnetic ordering with TN≈128T_N \approx128 K.
MnSe\mathrm{MnSe}Ref. [48]The high-pressure orthorhombic phase of MnSe shows AFM behavior.
NaCrS2\mathrm{NaCrS}_{2}Ref. [49]Antiferromagnetic ordering observed below about 18 K.
Pr2NiO4\mathrm{Pr}_{2}\mathrm{NiO}_{4}Ref. [50]Ni sublattice becomes three-dimensionally antiferromagnetically ordered at TN≈325T_N \approx325 K with propagation vector k=[100]\mathbf{k} = [100].
Rb2V3O8\mathrm{Rb}_{2}\mathrm{V}_{3}\mathrm{O}_{8}Ref. [51]Fresnoite-type vanadate with antiferromagnetic interactions below ∼10\sim10 K.
RbFe(MoO4)2\mathrm{RbFe(MoO}_{4}\mathrm{)}_{2}Ref. [52]Quasi-two-dimensional triangular-lattice antiferromagnet with noncollinear 120° order below TN≈3.8T_N \approx3.8 K.
VBr3\mathrm{VBr}_{3}Ref. [53]Van der Waals honeycomb antiferromagnet; neutron diffraction reveals magnetic Bragg peaks below TN=26.5T_N = 26.5 K with k=(0,1/2,1)\mathbf{k} = (0, 1/2, 1), corresponding to a zigzag AFM structure.
VF3\mathrm{VF}_{3}Ref. [54]NMR/susceptibility studies identify an antiferromagnetic transition at with TN≈19.4T_N \approx19.4 K.
YMn2O5\mathrm{YMn}_{2}\mathrm{O}_{5}Ref. [55]Antiferromagnetic domains observed directly by neutron spherical polarimetry.

Table III. Expanded literature audit of candidate non-ferromagnetic compounds from Supplementary Table S1 (intersection set between the XG Boost and Random Forest MAGNDATA classifiers) with references confirming their non-FM ordering.

Estimating the false discovery rate. Extending the naive propagation of the individual-classifier precisions (82% for RF and 81% for XGBoost) over the MAGNDATA validation dataset to the 7,843-compound intersection set gives false discovery rates of FDRRF_{\mathrm{RF}} = 1 − 0.82 = 18% and FDRXGB_{\mathrm{XGB}} = 1 − 0.81 = 19%, corresponding to approximately 1,400–1,500 expected false positives. This estimate, however, is a loose upper bound: it assumes that individual-classifier errors carry over unchanged to the intersection of two classifiers trained with different inductive biases, which overstates the true intersection-set FDR.

To quantify the false-discovery rate empirically of the intersection set (Supplementary Table S1), we drew a uniformly random sample of 100 compounds from Supplementary Table S1 and conducted a systematic literature audit of each (see Table VI). Of the 100 sampled compounds, zero were confirmed to be ferromagnetic. Four are confirmed non-FM by direct experiment on the specific compound (BaY2NiO5\mathrm{BaY}_{2}\mathrm{NiO}_{5}, Eu2MgSi2O7\mathrm{Eu}_{2}\mathrm{MgSi}_{2}\mathrm{O}_{7}, Mn2ZnO4\mathrm{Mn}_{2}\mathrm{ZnO}_{4}, TmMn2O5\mathrm{TmMn}_{2}\mathrm{O}_{5}), 22 are likely non-FM based on established behavior of their structural or chemical family, and 16 cannot be ferromagnetic because they contain no magnetic ions — their MP FM label is a pure DFT labeling artifact. Five compounds reside in chemical regimes where FM ordering could plausibly occur — two mixed-valence manganites supporting Zener double exchange (Ca2HoMn3O12\mathrm{Ca}_{2}\mathrm{HoMn}_{3}\mathrm{O}_{12}, SrPr2Mn3O9\mathrm{SrPr}_{2}\mathrm{Mn}_{3}\mathrm{O}_{9}), one Cr-chalcogenide spinel (Cr4InAgSe8\mathrm{Cr}_{4}\mathrm{InAgSe}_{8}) by analogy with the FM insulator family CdCr2Se4/HgCr2Se4\mathrm{CdCr}_{2}\mathrm{Se}_{4}/\mathrm{HgCr}_{2}\mathrm{Se}_{4}, one rare-

MaterialCat.ReferenceComments
Ba₂CaY(Co₄O₇)₂N—No direct magnetic data on this composition.
Ba₃Co₈Sn₃O₂₀N—No magnetic characterization found.
Ba₃Li(AsO₄)₂D—Ba²⁺, Li⁺, As⁵⁺ — all diamagnetic.
Ba₄Sr₂Nd₂Co₄O₁₅N—No magnetic characterization found.
BaLiGd₂(MoO₄)₄N—No magnetic characterization found.
BaN₆OD—No transition-metal or ff-electron ions.
BaNaCe₂C₄O₁₂FN—No magnetic data found.
BaPrEuSbO₆FRef. [74]Complex Pr/Eu-containing oxide; RR-bearing perovskites and double perovskites with Pr³⁺ consistently order AFM or remain paramagnetic at low TT, not FM.
BaSrGdSbO₆FRef. [74]Gd-based double perovskite; Gd³⁺ compounds typically exhibit AFM ordering at low TT.
BaY₂NiO₅CRef. [75]S=1S=1 one-dimensional Haldane-gap antiferromagnet; remains quantum disordered (no 3D order) down to 1.8 K.
CNClO₄D—No magnetic ions present.
Ca(ClO₂)₂D—Ca²⁺, Cl³⁺ — no unpaired electrons.
Ca(CoO₂)₂N—Co³⁺ layered oxide; no magnetic data found.
Ca₂HoMn₄BiO₁₂P—Quadruple-perovskite-type manganite. Mixed-valence manganites can host FM ordering via Zener double exchange. No direct data on this composition; FM cannot be ruled out.
Ca₂VP₃O₁₁N—V³⁺ phosphate; no magnetic data found.
CaMnCu₃(RuO₄)₃N—No study found.
Ce₂Te₄O₁₁N—Ce³⁺ tellurite; no magnetic study found.
CeY₄S₇N—Mixed-valent Ce sulfide; no magnetic study found.
Cr₃N₄FRef. [76]Cr nitride family; the well-characterized member CrN is AFM (TN≈280T_N \approx280 K). Cr₃N₄ has not been directly studied but is expected to show AFM Cr–N–Cr superexchange on similar grounds.
MaterialCat.ReferenceComments
Cr4InAgSe8\mathrm{Cr}_{4}\mathrm{InAgSe}_{8}P—Cr-chalcogenide spinel-like; Cr chalcogenide spinels (e.g., CdCr2Se4\mathrm{CdCr}_{2}\mathrm{Se}_{4}, HgCr2Se4\mathrm{HgCr}_{2}\mathrm{Se}_{4}) are well-known FM insulators with TC∼100T_C \sim100–130 K130\ \mathrm{K}, so FM order cannot be ruled out without direct data.
Cs3Cr2I9\mathrm{Cs}_{3}\mathrm{Cr}_{2}\mathrm{I}_{9}FRef. [77]Dimerized system with a singlet ground state of antiferromagnetically coupled Cr3+\mathrm{Cr}^{3+}–Cr3+\mathrm{Cr}^{3+} dimers. Spontaneous magnetic order at low TT arises from intratriplet condensation but is not simple FM.
Eu(ReO4)2\mathrm{Eu(ReO}_{4}\mathrm{)}_{2}N—No experimental data exist on this specific composition.
Eu2MgSi2O7\mathrm{Eu}_{2}\mathrm{MgSi}_{2}\mathrm{O}_{7}CRef. [78]151Eu{}^{151}\mathrm{Eu} Mössbauer confirms Eu2+\mathrm{Eu}^{2+}. Paramagnetic down to 1.8 K1.8\ \mathrm{K}; not FM.
FeTc2Ge\mathrm{FeTc}_{2}\mathrm{Ge}N—No data found.
Gd(Tl3Te2)3\mathrm{Gd(Tl}_{3}\mathrm{Te}_{2}\mathrm{)}_{3}P—Gd-based Zintl/intermetallic phase with conduction electrons and Gd3+\mathrm{Gd}^{3+} local moments; rare-earth intermetallics commonly host FM ordering via RKKY coupling. No study found, FM cannot be ruled out.
Gd2PbS4\mathrm{Gd}_{2}\mathrm{PbS}_{4}FRef. [74]Gd3+\mathrm{Gd}^{3+} (S=7/2S=7/2) sulfide; Gd3+\mathrm{Gd}^{3+} orders AFM at low TT in insulating hosts (cf. Gd2S3\mathrm{Gd}_{2}\mathrm{S}_{3}, TN≈1.5 KT_N \approx1.5\ \mathrm{K}).
HfCrAu2\mathrm{HfCrAu}_{2}N—Heusler-type intermetallic containing Cr; no magnetic data found.
HfMg14BO15\mathrm{HfMg}_{14}\mathrm{BO}_{15}D—Hf4+\mathrm{Hf}^{4+}, Mg2+\mathrm{Mg}^{2+}, B3+\mathrm{B}^{3+} — all diamagnetic.
HfMg14SiO16\mathrm{HfMg}_{14}\mathrm{SiO}_{16}D—Hf4+\mathrm{Hf}^{4+}, Mg2+\mathrm{Mg}^{2+}, Si4+\mathrm{Si}^{4+} — all diamagnetic.
HfMg6NbO8\mathrm{HfMg}_{6}\mathrm{NbO}_{8}D—Hf4+\mathrm{Hf}^{4+}, Mg2+\mathrm{Mg}^{2+}, Nb5+\mathrm{Nb}^{5+} (d0d^{0}) — no unpaired electrons.
K2Fe(PSe3)2\mathrm{K}_{2}\mathrm{Fe(PSe}_{3}\mathrm{)}_{2}FRef. [79]Iron phosphoselenide related to FePSe3_3 family; all members are Ising-type AFMs.
K2Fe2SeS3\mathrm{K}_{2}\mathrm{Fe}_{2}\mathrm{SeS}_{3}FRef. [79]Potassium iron mixed-chalcogenide; potassium iron chalcogenide family compounds consistently show AFM-type ordering, not FM.
K2Na2Mo(WO4)3\mathrm{K}_{2}\mathrm{Na}_{2}\mathrm{Mo(WO}_{4}\mathrm{)}_{3}N—No data found.
K4MgV6O16\mathrm{K}_{4}\mathrm{MgV}_{6}\mathrm{O}_{16}N—No magnetic data found.
KLiCoO2\mathrm{KLiCoO}_{2}N—No magnetic data found.
KMg3(SiO3)4\mathrm{KMg}_{3}\mathrm{(SiO}_{3}\mathrm{)}_{4}D—K+\mathrm{K}^{+}, Mg2+\mathrm{Mg}^{2+}, Si4+\mathrm{Si}^{4+} — all diamagnetic.
La2(SiIr)3\mathrm{La}_{2}\mathrm{(SiIr)}_{3}N—Ir intermetallic; no magnetic data found.
La3MnW(SO2)3\mathrm{La}_{3}\mathrm{MnW(SO}_{2}\mathrm{)}_{3}N—No study found.
LaWO2\mathrm{LaWO}_{2}N—No magnetic data found.
LaYNi10\mathrm{LaYNi}_{10}NRef. [80]Expected to be a Pauli paramagnet with no magnetic order, by analogy with its parent compounds LaNi5\mathrm{LaNi}_{5} and YNi5\mathrm{YNi}_{5}, but no study found.
Li2AgGe\mathrm{Li}_{2}\mathrm{AgGe}D—Li+\mathrm{Li}^{+}, Ag+\mathrm{Ag}^{+} (d10d^{10}), Ge — no unpaired electrons.
Li2MnV3O8\mathrm{Li}_{2}\mathrm{MnV}_{3}\mathrm{O}_{8}FRef. [21]Mixed Mn–V oxide; Mn2+\mathrm{Mn}^{2+} and V4+\mathrm{V}^{4+} exchange pathways are typically AFM, giving ferrimagnetic/AFM ordering rather than FM.
Li2VCO5\mathrm{Li}_{2}\mathrm{VCO}_{5}N—No study found.
Li3CrSi2O7\mathrm{Li}_{3}\mathrm{CrSi}_{2}\mathrm{O}_{7}N—Cr3+\mathrm{Cr}^{3+} silicate; no magnetic data found.
Li3Ti2V3O12\mathrm{Li}_{3}\mathrm{Ti}_{2}\mathrm{V}_{3}\mathrm{O}_{12}N—No magnetic data found.
Li3V(CO3)3\mathrm{Li}_{3}\mathrm{V(CO}_{3}\mathrm{)}_{3}N—V3+\mathrm{V}^{3+} carbonate; no study found.
Li4Cr(WO4)3\mathrm{Li}_{4}\mathrm{Cr(WO}_{4}\mathrm{)}_{3}FRef. [21]Paramagnetic to low TT with weak AFM coupling. No FM expected.
Li4Ni3WO8\mathrm{Li}_{4}\mathrm{Ni}_{3}\mathrm{WO}_{8}FRef. [81]Ni2+\mathrm{Ni}^{2+} (d8d^{8}) oxide; Ni2+\mathrm{Ni}^{2+} in octahedral oxide environments shows AFM superexchange (cf. La2NiO4\mathrm{La}_{2}\mathrm{NiO}_{4}, NiO).
Li6Cr5CuO12\mathrm{Li}_{6}\mathrm{Cr}_{5}\mathrm{CuO}_{12}FRef. [21]Cr3+\mathrm{Cr}^{3+}-rich oxide; Cr2O3\mathrm{Cr}_{2}\mathrm{O}_{3} is a prototypical AFM (TN=307 KT_N = 307\ \mathrm{K}).
Li6V3Sb3O16\mathrm{Li}_{6}\mathrm{V}_{3}\mathrm{Sb}_{3}\mathrm{O}_{16}N—No magnetic data found.
Li17Mn2(CoO4)3\mathrm{Li}_{17}\mathrm{Mn}_{2}\mathrm{(CoO}_{4}\mathrm{)}_{3}FRef. [21]Mn–Co lithium oxide; mixed Mn and Co in oxide environments typically give AFM/ferrimagnetic rather than FM ordering.
LiMgPb2\mathrm{LiMgPb}_{2}D—Li+\mathrm{Li}^{+}, Mg2+\mathrm{Mg}^{2+}, Pb2+\mathrm{Pb}^{2+} — all diamagnetic.
MaterialCat.ReferenceComments
LiMn3O5LiMn_{3}O_{5}FRef. [82]Li–Mn–O spinel family; LiMn2O4LiMn_{2}O_{4} is a frustrated AFM (ΘCW=−266\Theta_{\mathrm{CW}}=-266 K).
LiMnB2O5LiMnB_{2}O_{5}FRef. [21]Mn3+Mn^{3+} in borate framework; AFM Mn–O–Mn exchange where pathways exist, paramagnetic for isolated sites. FM not expected.
LiV2CrO6LiV_{2}CrO_{6}FRef. [21]V4+V^{4+} (d1d^{1}) and Cr3+Cr^{3+} (d3d^{3}) mixed-metal oxide; both give AFM superexchange.
LiVF5LiVF_{5}N—V4+V^{4+} fluoride; no magnetic data found.
Lu6MgWO12Lu_{6}MgWO_{12}D—Lu3+Lu^{3+} (4f144f^{14}), W6+W^{6+} (d0d^{0}) — no magnetic moment.
Mg(ClO5)2Mg(ClO_{5})_{2}D—Mg2+Mg^{2+}, Cl — no unpaired electrons.
Mg14CoCO16Mg_{14}CoCO_{16}N—No data found.
Mg14TiCrO16Mg_{14}TiCrO_{16}N—No data found.
Mg30VCuO32Mg_{30}VCuO_{32}N—No data found.
Mg30VZnO32Mg_{30}VZnO_{32}N—No data found.
MgCo2(BO3)2MgCo_{2}(BO_{3})_{2}N—No magnetic data found.
MgMn(SbO3)4MgMn(SbO_{3})_{4}N—Mn2+Mn^{2+} antimonate; no study found.
MgMn2(SiO3)4MgMn_{2}(SiO_{3})_{4}N—Mn3+Mn^{3+} silicate; no magnetic data found.
MgMn2Cr2O8MgMn_{2}Cr_{2}O_{8}FRef. [83, 84, 85]Mixed Mn–Cr oxide; AFM superexchange between Mn and Cr magnetic sublattices is expected to give ferrimagnetic or AFM (non-FM) ordering.
MgV2Bi4O11MgV_{2}Bi_{4}O_{11}N—V4+V^{4+} bismuth vanadate; no study found.
Mn(CuCl2)2Mn(CuCl_{2})_{2}N—Isolated Mn2+Mn^{2+} in diamagnetic Cu+Cu^{+} (d10d^{10}) chloride framework; no magnetic data found.
Mn2ZnO4Mn_{2}ZnO_{4}CRef. [86]Magnetic measurements at 2 K confirm the formation of antiferromagnetic order.
MnSb6(Pb2S7)2MnSb_{6}(Pb_{2}S_{7})_{2}N—Mn2+Mn^{2+} sulfosalt; no study found.
MnZnCdTe3MnZnCdTe_{3}N—No data found.
Na2Mg(CoO2)2Na_{2}Mg(CoO_{2})_{2}FRef. [21]Co2+Co^{2+}-containing layered oxide; Co2+Co^{2+} in octahedral oxide environments shows strong AFM superexchange (cf. CoO, TN=291T_{N}=291 K). Not expected to be FM.
NaCr3O8NaCr_{3}O_{8}PRef. [87]Cr5+Cr^{5+} (d1d^{1}) chromate; DFT-predicted to be a FM insulator. No conclusive experimental confirmation of long-range FM order.
NaMg14FeO16NaMg_{14}FeO_{16}N—No data found.
NaMnCuSe2NaMnCuSe_{2}N—Isolated Mn2+Mn^{2+} in diamagnetic Cu+Cu^{+} selenide framework; no magnetic data found.
NaNdMnWO6NaNdMnWO_{6}FRef. [21]Ordered double perovskite; Mn and W alternate on the B-site, giving long-range Mn–O–W–O–Mn super-superexchange. Mn-based ordered double perovskites consistently order AFM at low TT.
NaNi2Mo2H3O10NaNi_{2}Mo_{2}H_{3}O_{10}N—Ni2+Ni^{2+} molybdate hydrate; no data found.
PtN4(Cl2O)2PtN_{4}(Cl_{2}O)_{2}D—Pt4+Pt^{4+} (d6d^{6} low-spin); typically diamagnetic.
Rb2V2O5Rb_{2}V_{2}O_{5}N—V4+V^{4+} (d1d^{1}) vanadate; no magnetic data found.
Rb2V4O9Rb_{2}V_{4}O_{9}N—V4+V^{4+} (d1d^{1}) vanadate; no magnetic data found.
Rb4CO6Rb_{4}CO_{6}N—No data found.
RbV2SeO7RbV_{2}SeO_{7}N—Mixed-valence V selenite; no study found.
Sc2GaHgSc_{2}GaHgD—Sc, Ga, Hg intermetallic; no local moments expected.
ScMn2O5ScMn_{2}O_{5}FRef. [88]Mullite-type RMn2O5RMn_{2}O_{5} family (studied RR: Y, Tb–Lu) orders AFM with incommensurate k; Sc member expected to follow family behavior.
Sr2AlTlCo2O7Sr_{2}AlTlCo_{2}O_{7}N—No magnetic data found.
Sr3La(NiO4)2Sr_{3}La(NiO_{4})_{2}FRef. [81]Ruddlesden–Popper-type nickelate with mixed-valence Ni; related La–Ni–O RP phases (La2NiO4La_{2}NiO_{4}, La3Ni2O7La_{3}Ni_{2}O_{7}) show AFM or spin-density-wave ordering, not FM.
Sr3MgV2O7Sr_{3}MgV_{2}O_{7}N—V3+V^{3+} double perovskite; no data found.
SrNdMgRuO6SrNdMgRuO_{6}FRef. [89]By analogy with the Sr2LnRuO6Sr_{2}LnRuO_{6} family, SrNdMgRuO6SrNdMgRuO_{6} is expected to develop antiferromagnetic order.
MaterialCat.ReferenceComments
SrPr2_2Mn3_3O9_9P—Ruddlesden–Popper-type manganite with mixed-valence Mn3+^{3+}/Mn4+^{4+}; such mixed-valence manganites can host FM ordering via Zener double exchange. No direct data on this composition; FM cannot be ruled out.
TcCl4_4N—Tc4+^{4+} (d0d^0) chloride; Tc is radioactive. No data.
Ti2_2AlWD—Ti, Al, W intermetallic; Ti-based intermetallics typically Pauli paramagnetic, no local moments.
TmMn2_2O5_5CRef. [90]AFM multiferroic with an ordering temperature of ∼40\sim40 K.
U(Al2_2Cu)4_4N—U intermetallic; no magnetic data found.
U4_4TeSe7_7N—U chalcogenide; no magnetic data found.
UGeI6_6N—U4+^{4+} halide; no magnetic data found.
V5_5Zn4_4(TeO6_6)3_3N—V tellurate; no study found.
VP2_2Pb2_2O9_9N—V4+^{4+} phosphate; no study found.
Y(VS2_2)2_2N—V sulfide (Y3+^{3+} diamagnetic); no data found.
YNbO2_2N—No experimental data.
ZnBP2_2PbO9_9D—Zn2+^{2+}, B3+^{3+}, P5+^{5+}, Pb2+^{2+} — all diamagnetic.

Table VI. TABLE VI: Expanded literature audit of compounds from Supplementary Table S1 (intersection set), based on a random sample of 100 materials (seed = 42). Categories: C = confirmed non-FM by direct experiment (4); F = likely non-FM from family-level evidence (22); D = no magnetic ions, FM label is a DFT artifact (16); P = potential FM order that cannot be ruled out (e.g. mixed-valence manganites supporting double exchange, Cr-chalcogenide spinels, or rare-earth intermetallics with RKKY coupling) (5); N = no experimental magnetic data found (53). Compounds in the P category are conservatively treated as potential false positives; see the Estimating the false discovery rate part of Section 3.

earth intermetallic with RKKY coupling (Gd(Tl3_3Te2_2)3_3), and one DFT-predicted FM insulator (NaCr3_3O8_8) — and are conservatively classified as potential false positives. The remaining 53 compounds have no published experimental magnetic data with which to compare directly. Even under the most conservative accounting — treating all five potential-FM candidates as genuine false positives — the empirical FDR on this random sample is bounded at ≲5%\lesssim5\%, less than a third of the naive 18–19% ceiling obtained by propagating individual-classifier precisions. This validates the premise that taking the intersection of two classifiers with different inductive biases substantially suppresses error rates below either model alone.

Removing noisy entries from the candidate material datasets. Finally, direct screening of the full intersection set shows that 1,003 of the 7,843 candidates (12.8%) contain no magnetic ions at all, confirming that a substantial fraction of MP’s FM labels are pure DFT labeling artifacts — consistent and close to the 16% ”no magnetic ion” rate observed in our random-sample audit (Table VI). Candidate non-FM materials with at least one magnetic atom predicted by each classifier have a similar fraction of FM artifacts and are also listed in the Supplementary Tables 4-6, where the number of XGBoost flagged candidates reduce from 9,383 to 8,189 (12.7% are DFT artifacts), Random Forest from 10,354 to 9,055 (12.5% are DFT artifacts), and their intersection set reduces from 7,843 to 6,840. Although our classifiers correctly identify these compounds as having nonzero propagation vector (i.e., not simple FM), flagging such compounds as candidates for being FM-mislabeled is trivially correct. Nonetheless, we retain the D-category entries are retained in Table VI because they independently show that the classifiers are still capable of identifying these compounds.

Effect of additional extra physically motivated descriptors. Having established the performance of the classifiers using our minimal set of descriptors, we now relax the restriction described in Section II and re-train the classifiers with two additional physically motivated compositional descriptors: the composition-weighted mean Pauling electronegativity and the composition-weighted mean atomic dipole polarizability. These descriptors are related to screened Coulomb interactions and have been used as predictors in a recent machine-learning study of magnetic ordering [17]. Per-element electronegativities were obtained from pymatgen [56], and per-element polarizabilities from the mendeleev Python package [57], which sources its values from the CRC Handbook of Chemistry and Physics [58] and from Ref. [59].

The effect on the performance of our best classifiers is summarized in Table V and IV. The cross-validated F1macroF_1^{\mathrm{macro}} improves by approximately 1% on the four-class MP task across the different classifiers (e.g. from 0.658 to 0.665 for LightGBM) and by approximately 1–2% on the binary MAGNDATA propagation-vector task (e.g. from 0.883 to 0.906 for Random Forest). The improvement is more concentrated on the minority classes, as expected when more discriminative features benefit the harder-to-distinguish categories more. The modest overall increase in the classifiers efficiency indicates that the tree-ensemble classifiers could already implicitly capture much of the compositional-averaging information that explicit means of unseen descriptors encode. A more substantial gain would likely require local structural features — coordination environments, magnetic-site connectivity, or other descriptors that global compositional means cannot encode.

Class\multicolumn{3}{c}{Without Xˉ\bar{X}, αˉ\bar{\alpha}}\multicolumn{3}{c}{With Xˉ\bar{X}, αˉ\bar{\alpha}}
Prec.Rec.F1F_1Prec.Rec.F1F_1
AFM0.420.480.450.450.460.46
FM0.850.820.840.850.850.85
FiM0.410.470.440.450.490.47
NM0.940.930.930.940.930.94
Accuracy0.820.83
Macro avg0.650.670.660.670.680.68
Weighted avg0.830.820.820.840.830.83
Mean CV F1macroF_1^{\mathrm{macro}}0.6580.665

Table V. Effect of adding composition-weighted mean Pauling electronegativity (Xˉ\bar{X}) and atomic dipole polarizability (αˉ\bar{\alpha}) on the LightGBM classifier performance for the four-class MP magnetic-ordering task. Per-class precision, recall, and F1F_1 are reported on the validation set. CV F1F_1-macro is the mean over 5 stratified folds on the training set.

ClassWithout Xˉ\bar{X}, αˉ\bar{\alpha} Prec.Without Xˉ\bar{X}, αˉ\bar{\alpha} Rec.Without Xˉ\bar{X}, αˉ\bar{\alpha} F1F_1With Xˉ\bar{X}, αˉ\bar{\alpha} Prec.With Xˉ\bar{X}, αˉ\bar{\alpha} Rec.With Xˉ\bar{X}, αˉ\bar{\alpha} F1F_1
nonzero0.820.900.860.840.900.87
zero0.970.940.950.970.950.96
Accuracy0.930.94
Macro avg0.900.920.910.910.920.91
Weighted avg0.930.930.930.940.940.94
Mean CV F1macroF_1^{\mathrm{macro}}0.8830.906

Table 4. Effect of adding composition-weighted mean Pauling electronegativity (Xˉ\bar{X}) and atomic dipole polarizability (αˉ\bar{\alpha}) on the Random Forest classifier performance for the binary propagation-vector classification task on MAGNDATA. Per-class precision, recall, and F1F_1 are reported on the validation set. CV F1F_1-macro is the mean over 6 stratified folds on the training set.

Conclusion

In conclusion, a central outcome of this work is the identification of a systematic ferromagnetic bias in the Materials Project database, revealed through our propagation-vector classifiers trained on experimentally validated MAGNDATA entries. While Materials Project magnetic labels frequently default to FM orderings stemming from the use of ferromagnetic initialization in automated DFT workflows, even when the true ground state is antiferromagnetic, ferrimagnetic, or complex, our models consistently flagged these discrepancies for more than 6,840 candidate compounds, with the empirical false-discovery rate estimated by a random-sample validation to be ≲5%\lesssim5\%. Specifically, the Random Forest and XGBoost classifiers we developed achieved accuracies of over 92% (with F1F_1-macro scores above 90%) in distinguishing zero from nonzero propagation vector structures. Importantly, our approach demonstrates how machine learning can serve not only as a predictive tool but also as a diagnostic layer to expose and correct database-level artifacts, offering a scalable pathway to improve the reliability of high-throughput materials databases.

Limitations. However, our approach has several limitations that should be kept in mind. First, the class distribution in MAGNDATA is strongly unbalanced: antiferromagnetic entries dominate the database, with ferromagnetic, and ferrimagnetic compounds much more sparsely represented. As a result, a classifier trained directly on MAGNDATA cannot reliably learn to distinguish all four magnetic classes simultaneously, and this is the main reason we opted for a binary propagation-vector classifier (zero versus nonzero) rather than a four-class magnetic-order classifier on MAGNDATA. Second, although the tree-based architectures used in this work are efficient in classification tasks [60], they are still insufficient to predict the full three-dimensional magnetic structure of a given compound. Our classifiers answer coarse-grained binary or four-way questions, but reconstructing the complete spin configuration — including the direction of the ordered moments, the detailed periodicity of modulated phases, and non-collinear arrangements — requires more expressive architectures such as graph neural networks that operate directly on atomic coordinates and local environments. Third, our feature set is deliberately minimal, consisting only of elemental composition, and a handful of global electronic and structural quantities available from the Materials Project API. Richer descriptors — including coordination environments, bond-length distributions, magnetic-site connectivity graphs, and explicit spin-orbit or Hubbard U estimates — would likely improve both accuracy and physical interpretability, and could allow the classifiers to resolve finer distinctions such as collinear versus spiral antiferromagnetism that are currently merged into the single nonzero propagation vector class.

Future Directions. Now we suggest several concrete directions as natural next steps to build on this work. Perhaps the most immediate is to move beyond the tree-ensemble, global-descriptor approach and adopt architectures that operate directly on atomic structure, such as crystal graph convolutional networks and Euclidean equivariant neural networks [17, 61], which preserve crystallographic symmetry by construction and can in principle predict the full three-dimensional magnetic structure rather than the binary summary we consider here. More recent spin-dependent and time-reversal equivariant networks [62] go further and explicitly incorporate the spin degree of freedom into the network architecture, allowing direct prediction of non-collinear configurations and spin-lattice coupling. Additionally, some materials exhibit more than one magnetic structure depending at the temperature, therefore incorporating temperature effects on the formation of the magnetic order is a higher level future task. A second direction is to enlarge and rebalance the training data: the recently compiled Northeast Materials Database (NEMAD), which contains about 67,500 experimental-based reported magnetic entries extracted from the literature using large language models [63], offers a substantially larger and more class-balanced training set than MAGNDATA, and incorporating it would partially address the class-imbalance limitation we discuss above. Where balancing the database itself is not possible, data-level interventions such as SMOTE-style oversampling [64] and loss-level cost-sensitive learning [65] are both straightforward to integrate into the existing pipeline. A third direction is to close the loop between the machine learning screen and the underlying abab initio calculations: our flagged candidate list can serve as input to an automated DFT+UU workflow with explicit AFM enumeration in the spirit of Horton et al. [29], which would convert our heuristic flags into concrete proposed corrections to the MP labels and, through active-learning-style feedback [66], progressively refine the classifiers themselves. Fourth, targeted experimental follow-up on high-interest compounds from the flagged list — particularly those whose chemistry is unusual enough that no obvious family-level analog exists — remains the most decisive test of the approach, and the one most likely to uncover possible new magnetic families.

A fifth direction, more speculative but physically well-motivated, connects our approach to the rapidly developing field of magnetic topological materials. Topological quantum chemistry and symmetry-indicator methods [67, 68, 69] classify the topological phase of a crystalline material from the symmetry data of its occupied electronic bands, and the extension of this framework to magnetic crystals [70, 71, 72] uses the magnetic space group — which depends on the full magnetic structure, including propagation vector and spin orientation — as the relevant symmetry input. If future machine learning or deep learning architectures of the kind discussed above could ultimately predict the full three-dimensional magnetic structure of a compound from its chemistry and structure alone, their output can be fed directly into this magnetic-topological-quantum-chemistry pipeline, enabling large-scale screening of magnetic topological insulators, semimetals, axion insulators, and topological magnon compounds [73] without the prohibitive cost of performing a full DFT plus symmetry analysis on every candidate material. MAGNDATA is an especially natural starting point for this integration, since its entries already specify the full magnetic structure from which the magnetic space group can be determined unambiguously.

Acknowledgments

This research was primarily supported by the Center for Emergent Materials, an NSF MRSEC, under award number DMR−2011876. We thank Murod Mirzhalilov, Brandon Abrego, and Steven Gubkin for helpful discussions.

Data Availability

The supplementary tables of flagged candidate compounds, including the full intersection set and the per-classifier candidate lists are openly available in machine-readable CSV format on Zenodo at https://zenodo.org/records/20044097. The Materials Project data used in this work were obtained from the public Materials Project API (https://materialsproject.org); the MAGNDATA labels were obtained from the Bilbao Crystallographic Server (https://cryst.ehu.es/index.html).

Appendix

Supplementary Note

All stochastic components of the pipeline — data splitting, cross-validation fold assignment, hyperparameter search sampling, and per-model random operations (bootstrap, feature subsampling, etc.) — are controlled by a single fixed seed, random_state = 42, ensuring that the reported results are fully reproducible.

References

References

  1. [1]A. J. Cohen, P. Mori-Sánchez, and W. Yang, Insights into current limitations of density functional theory, Science 321, 792 (2008).DOI
  2. [2]B. Himmetoglu, A. Floris, S. de Gironcoli, and M. Cococcioni, Hubbard-corrected dft energy functionals: The lda+u description of correlated systems, International Journal of Quantum Chemistry 114, 14 (2014).DOI
  3. [3]A. Jain, J. Montoya, S. Dwaraknath, N. E. R. Zimmermann, J. Dagdelen, M. Horton, P. Huck, D. Winston, S. Cholia, S. P. Ong, and K. A. Persson, The materials project: Accelerating materials design through theory-driven data and tools, in Handbook of Materials Modeling (Springer, 2018) pp. 1–34.
  4. [4]G. Kotliar, S. Y. Savrasov, K. Haule, V. S. Oudovenko, O. Parcollet, and C. A. Marianetti, Electronic structure calculations with dynamical mean-field theory, Reviews of Modern Physics 78, 865 (2006).DOI
  5. [5]W. M. C. Foulkes, L. Mitas, R. J. Needs, and G. Rajagopal, Quantum monte carlo simulations of solids, Reviews of Modern Physics 73, 33 (2001).
  6. [6]W. E. Pickett, Pseudopotential methods in condensed matter applications, Physical Review B 39, 4814 (1989).DOI
  7. [7]O. K. Andersen, Linear methods in band theory, Physical Review B 12, 3060 (1975).
  8. [8]S. W. Lovesey, Theory of Neutron Scattering from Condensed Matter (Oxford University Press, 1986).
  9. [9]I. A. Zaliznyak and J. M. Tranquada, Neutron scattering and its application to strongly correlated systems, in Springer Series in Solid-State Sciences (Springer, 2012) pp. 1–64, available online: https://qss3.mit.edu/images/pdf/neutrons_v3.pdf.DOI
  10. [10]L. J. P. Ament, M. van Veenendaal, T. P. Devereaux, J. P. Hill, and J. van den Brink, Resonant inelastic x-ray scattering studies of elementary excitations, Reviews of Modern Physics 83, 705 (2011).DOI
  11. [11]S. J. Blundell, Spin-polarized muons in condensed matter physics, Contemporary Physics 40, 175 (1999).DOI
  12. [12]S. V. Gallego, J. M. Perez-Mato, L. Elcoro, E. S. Tasci, R. M. Hanson, and M. I. Aroyo, Magndata: Towards a database of magnetic structures. i. the commensurate case, Journal of Applied Crystallography 49, 1750 (2016).DOI
  13. [13]S. V. Gallego, J. M. Perez-Mato, M. I. Aroyo, E. S. Tasci, D. Orobengoa, L. Elcoro, K. Momma, and G. Madariaga, Magndata: Towards a database of magnetic structures. ii. the incommensurate case, Journal of Applied Crystallography 49, 1941 (2016).DOI
  14. [14]V. F. Sears, Neutron scattering lengths and cross sections, Neutron News 3, 26 (1992).
  15. [15]M. Eriksson, J. F. van der Veen, and C. Quitmann, Diffraction-limited storage rings – a window to the science of tomorrow, Journal of Synchrotron Radiation 21, 837 (2014).DOI
  16. [16]A. Ghosh, F. Ronning, S. M. Nakhmanson, and J.-X. Zhu, Machine learning study of magnetism in uranium-based compounds, Physical Review Materials 4, 064414 (2020).DOI
  17. [17]H. A. Merker, H. Heiberger, L. Nguyen, T. Liu, Z. Chen, N. Andrejevic, N. C. Drucker, R. Okabe, S. E. Kim, Y. Wang, T. Smidt, and M. Li, Machine learning magnetism classifiers from atomic coordinates, iScience 25, 105192 (2022).DOI
  18. [18]J. Rodríguez-Carvajal and J. Villain, Magnetic structures, Comptes Rendus Physique 20, 770 (2019).
  19. [19]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, Commentary: The Materials Project: A materials genome approach to accelerating materials innovation, APL Materials 1, 011002 (2013).DOI
  20. [20]P. W. Anderson, Antiferromagnetism. theory of superexchange interaction, Phys. Rev. 79, 350 (1950).DOI
  21. [21]J. B. Goodenough, Magnetism and the Chemical Bond (Wiley-Interscience, New York, 1963).
  22. [22]J. Kanamori, Superexchange interaction and symmetry properties of electron orbitals, J. Phys. Chem. Solids 10, 87 (1959).DOI
  23. [23]E. C. Stoner, Collective electron ferromagnetism, Proc. R. Soc. Lond. A 165, 372 (1938).
  24. [24]J. Kübler, Theory of Itinerant Electron Magnetism, 2nd ed., International Series of Monographs on Physics (Oxford University Press, Oxford, 2021).
  25. [25]A. W. Overhauser, Spin density waves in an electron gas, Phys. Rev. 128, 1437 (1962).
  26. [26]E. Fawcett, Spin-density-wave antiferromagnetism in chromium, Rev. Mod. Phys. 60, 209 (1988).DOI
  27. [27]N. F. Mott, Metal-Insulator Transitions (Taylor & Francis, London, 1990).
  28. [28]P. Coleman, Heavy fermions: Electrons at the edge of magnetism, in Handbook of Magnetism and Advanced Magnetic Materials (Wiley, 2007).arxiv.org/abs/cond-mat/0612006
  29. [29]M. K. Horton, J. H. Montoya, M. Liu, and K. A. Persson, High-throughput prediction of the ground-state collinear magnetic order of inorganic materials using density functional theory, npj Computational Materials 5, 64 (2019).DOI
  30. [30]D. M. W. Powers, Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation, Journal of Machine Learning Technologies 2, 37 (2011).
  31. [31]M. Sokolova and G. Lapalme, A systematic analysis of performance measures for classification tasks, Information Processing & Management 45, 427 (2009).DOI
  32. [32]F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, A. Müller, J. Nothman, G. Louppe, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and Édouard Duchesnay, Scikit-learn: Machine learning in python (2018), arXiv:1201.0490 [cs.LG].
  33. [33]Y. He, P.-P. De Breuck, H. Weng, M. Giantomassi, and G.-M. Rignanese, Machine learning on multiple topological materials datasets, npj Computational Materials 11, 181 (2025).arxiv.org/abs/2503.16276
  34. [34]F. Berardi, L. A. V. Nagle-Cocco, J. M. A. Steele, X. Zhang, C. Liu, H. E. Fischer, and S. E. Dutton, Structural phase transitions and magnetic characterization of Ba2GdNbO6 for low-temperature magnetocaloric refrigeration, Chemistry of Materials 37, 8848 (2025).DOI
  35. [35]E. C. Koskelo, C. Liu, P. Mukherjee, N. D. Kelly, and S. E. Dutton, Free-spin dominated magnetocaloric effect in dense gd3+ double perovskites (2022), arXiv:2203.08214 [cond-mat.str-el].arxiv.org/abs/2203.08214
  36. [36]Y. Doi and Y. Hinatsu, Magnetic properties of ordered perovskites Ba2LnTaO6 (Ln = Y, lanthanides), Journal of Physics: Condensed Matter 13, 4191 (2001).DOI
  37. [37]K. Sano and K. Takano, Spin gap of two-dimensional antiferromagnet representing CaV4O9, Journal of the Physical Society of Japan 65, 46 (1996).
  38. [38]M. Leblanc, D. Grebille, Y. Piffard, W. Massa, and O. V. Yakubovich, The magnetic structures of Mn2PO4F and Co2PO4F at 1.2 k, Journal of Magnetism and Magnetic Materials 173, 97 (1997).DOI
  39. [39]SpringerMaterials, CsCr(MoO4)2 curie-weiss paramagnetism, SpringerMaterials, PAULING FILE database (2024), classification by properties: Curie-Weiss paramagnetism; accessed 2026-04-12.
  40. [40]J. J. B. Blake, A. A. Labh, V. Pomjakushin, U. Keiderling, A. C. Komarek, L. Zhao, J. Baas, C. Pappas, and G. R. Blake, Magnetic properties of a non-centrosymmetric polymorph of fecl3, Materials Advances 6, 4471 (2025).
  41. [41]S. Ebisu, Y. Iijima, T. Iwasa, and S. Nagata, Antiferromagnetic transition and electrical conductivity in α-gd2s3, Journal of Physics and Chemistry of Solids 65, 1113 (2004).DOI
  42. [42]K. Somesh, Y. Furukawa, G. Simutis, F. Bert, M. Prinz-Zwick, N. Büttgen, A. Zorko, A. A. Tsirlin, P. Mendels, and R. Nath, Universal fluctuating regime in triangular chromate antiferromagnets, Physical Review B 104, 104422 (2021).DOI
  43. [43]B. van Laar and F. Engelsman, The magnetic structure of KCrS2, Journal of Solid State Chemistry 6, 384 (1973).DOI
  44. [44]A. I. Smirnov, L. E. Svir’tov, L. A. Prozorova, A. Zheludev, M. D. Lumsden, E. Ressouche, O. A. Petrenko, M. Hagiwara, S. Kimura, K. Nishikawa, K. Kindo, and A. Y. Shapiro, Coexistence of spiral and commensurate structures in a triangular antiferromagnet kfe(moo4)2, Journal of Physics: Conference Series 200, 032068 (2010).DOI
  45. [45]SpringerMaterials, LaCrTeO6 (LaCrTeO6) crystal structure, SpringerMaterials, PAULING FILE database (2024), classification by properties: antiferromagnet (AFM); accessed 2026-04-12.
  46. [46]J. Cong, K. Zhai, Y. Chai, D. Shang, D. D. Khalyavin, R. D. Johnson, D. P. Kozlenko, S. E. Kichanov, A. M. Abakumov, A. Tsirlin, L. D. Browinsky, X. Xu, Z. Sheng, S. V. Ovsyannikov, and Y. Sun, Spin-induced multiferroicity in the binary perovskite manganite mn2o3, Nature Communications 9, 2996 (2018).DOI
  47. [47]A. Punnoose, H. Magnone, and M. S. Seehra, Synthesis and antiferromagnetism of mn5o8, IEEE Transactions on Magnetics 37, 2150 (2001).
  48. [48]X.-X. Ma, B.-C. Gong, P.-H. Sun, K. Liu, and Z.-Y. Lu, First-principles study on the magnetic and electronic properties of the high-pressure orthorhombic phase of mnse, Physical Review B 106, 035136 (2022).DOI
  49. [49]K. W. Blazey and H. Rohrer, Antiferromagnetic phase diagram and magnetic band gap shift of NaCrS2, Physical Review 185, 712 (1969).DOI
  50. [50]M. T. Fernández-Díaz, J. Rodríguez-Carvajal, J. L. Martínez, G. Fillion, F. Fernández, and R. Sáez-Puche, Structural and magnetic phase transitions in Pr₂NiO₄, Zeitschrift für Physik B Condensed Matter 82, 275 (1991).DOI
  51. [51]G. Liu and J. E. Greedan, Magnetic properties of fresnoite-type vanadium oxides: A₂V₃O₈ (A = K, Rb, NH₄), Journal of Solid State Chemistry 114, 499 (1995).DOI
  52. [52]L. E. Svistov, A. I. Smirnov, L. A. Prozorova, O. A. Petrenko, L. N. Demianets, and A. Y. Shapiro, Quasi-two-dimensional antiferromagnet on a triangular lattice rbfe(moo₄)₂, Physical Review B 67, 094434 (2003).DOI
  53. [53]Y. Gu, Y. Hao, Z. Kao, Y. Gu, F. Liu, S. Zheng, H. Cao, L. He, and J. Zhao, Lattice and magnetic structure in the van der waals antiferromagnet vbr₃, Physical Review B 110, 064403 (2024).DOI
  54. [54]A. C. Gossard, K. L. Guggenheim, F. S. L. Hsu, and R. C. Sherwood, Magnetic ordering of vf₃, in AIP Conference Proceedings, Vol. 5 (1972) pp. 302–306.DOI
  55. [55]P. G. Radaelli, L. C. Chapon, A. Daoud-Aladine, C. Vecchini, P. J. Brown, T. Chatterji, S. Park, and S.-W. Cheong, Electric field switching of antiferromagnetic domains in ymn₂o₅: A probe of the multiferroic mechanism, Physical Review Letters 101, 067205 (2008).DOI
  56. [56]S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Python materials genomics (pymatgen): A robust, open-source python library for materials analysis, Computational Materials Science 68, 314 (2013).DOI
  57. [57]L. Mentel, mendeleev – a Python resource for properties of chemical elements, ions and isotopes, https://github.com/lmmentel/mendeleev (2014).
  58. [58]J. R. Rumble, ed., CRC Handbook of Chemistry and Physics, 104th ed. (CRC Press, Boca Raton, FL, 2023).
  59. [59]P. Schwerdtfeger and J. K. Nagy, 2018 table of static dipole polarizabilities of the neutral elements in the periodic table, Molecular Physics 117, 1200 (2019).
  60. [60]L. Grinsztajn, E. Oyallion, and G. Varoquaux, Why do tree-based models still outperform deep learning on tabular data? (2022), arXiv:2207.08815 [cs.LG].arxiv.org/abs/2207.08815
  61. [61]K. Choudhary, P. Minch, T. D. Rhone, and R. Bhattarai, Prediction of magnetic properties in van der waals magnets using graph neural networks, Physical Review Materials 8, 114002 (2024).DOI
  62. [62]H. Yu, Y. Zhong, L. Hong, C. Xu, W. Ren, X. Gong, and H. Xiang, Spin-dependent graph neural network potential for magnetic materials, Physical Review B 109, 144426 (2024).DOI
  63. [63]S. S. Itani et al., The northeast materials database for magnetic materials, Nature Communications 16, 9345 (2025), exact author list and article number: please verify from the published version.arxiv.org/abs/2409.15675
  64. [64]N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, SMOTE: Synthetic minority over-sampling technique, Journal of Artificial Intelligence Research 16, 321–357 (2002).DOI
  65. [65]S. H. Khan, M. Hayat, M. Bennamoun, F. Sohel, and R. Togneri, Cost sensitive learning of deep feature representations from imbalanced data (2017), arXiv:1508.03422 [cs.CV].arxiv.org/abs/1508.03422
  66. [66]T. Long, N. M. Fortunato, I. Opahle, Y. Zhang, I. Samathrakis, C. Shen, O. Gutfeisch, and H. Zhang, Accelerating the discovery of novel magnetic materials using machine learning-guided adaptive feedback, Proceedings of the National Academy of Sciences 119, e2204485119 (2022).DOI
  67. [67]B. Bradlyn, L. Elcoro, J. Cano, M. G. Vergniory, Z. Wang, C. Felser, M. I. Aroyo, and B. A. Bernevig, Topological quantum chemistry, Nature 547, 298 (2017).DOI
  68. [68]H. C. Po, A. Vishwanath, and H. Watanabe, Symmetry-based indicators of band topology in the 230 space groups, Nature Communications 8, 50 (2017).DOI
  69. [69]J. Kruthoff, J. de Boer, J. van Wezel, C. L. Kane, and R.-J. Slager, Topological classification of crystalline insulators through band structure combinatorics, Physical Review X 7, 041069 (2017).DOI
  70. [70]H. Watanabe, H. C. Po, and A. Vishwanath, Structure and topology of band structures in the 1651 magnetic space groups, Science Advances 4, eaat8685 (2018).DOI
  71. [71]Y. Xu, L. Elcoro, Z.-D. Song, B. J. Wieder, M. G. Vergniory, N. Regnault, Y. Chen, C. Felser, and B. A. Bernevig, High-throughput calculation of magnetic topological materials, Nature 586, 702 (2020).DOI
  72. [72]L. Elcoro, B. J. Wieder, Z. Song, Y. Xu, B. Bradlyn, and B. A. Bernevig, Magnetic topological quantum chemistry, Nature Communications 12, 5965 (2021).DOI
  73. [73]M. J. Karaki, A. E. Fahmy, A. J. Williams, S. Haravifard, J. E. Goldberger, and Y.-M. Lu, High-throughput discovery of perturbation-induced topological magnons, npj Computational Materials 11, 10.1038/s41524-025-01706-2 (2025).DOI
  74. [74]J. Jensen and A. R. Mackintosh, Rare Earth Magnetism: Structures and Excitations (Clarendon Press, Oxford, 1991).
  75. [75]G. Xu, J. F. DiTusa, T. Ito, K. Oka, H. Takagi, C. Broholm, and G. Aeppli, ybani₀₅: A nearly ideal realization of the s = 1 heisenberg chain with antiferromagnetic interactions, Phys. Rev. B 54, R6827(R) (1996).pubmed.ncbi.nlm.nih.gov/9984390
  76. [76]L. M. Corliss, N. Elliott, and J. M. Hastings, Antiferromagnetic structure of CrN, Phys. Rev. 117, 929 (1960).DOI
  77. [77]B. Leuenberger, H. U. Güdel, and P. Fischer, Synthesis, structural characterization, and magnetic properties of the dimer compounds Cs₃Cr₂X₉, X = Cl, Br, I, J. Solid State Chem. 64, 80 (1986).DOI
  78. [78]T. Endo, Y. Doi, M. Wakeshima, and Y. Hinatsu, Crystal structures and magnetic properties of new europium melilites Eu₂MSi₂O₇ (M = Mg, Mn) and their strontium analogues, Inorg. Chem. 49, 10809 (2010).DOI
  79. [79]A. Wiedenmann, J. Rossat-Mignod, A. Louisy, R. Brec, and J. Rouxel, Neutron diffraction study of the layered compounds MnPSe₃ and FePSe₃, Solid State Commun. 40, 1067 (1981).DOI
  80. [80]G. E. Grechnev, A. V. Logosha, A. S. Panfilov, A. G. Kuchin, and A. N. Vasiliev, Effect of pressure on the magnetic properties of YNi₅, LaNi₅, and CeNi₅, Low Temperature Physics 37, 138 (2011).DOI
  81. [81]J. Rodríguez-Carvajal, M. T. Fernández-Díaz, and J. L. Martínez, Neutron diffraction study on structural and magnetic properties of La₂NiO₄, J. Phys.: Condens. Matter 3, 3215 (1991).
  82. [82]A. S. Wills, N. P. Raju, and J. E. Greedan, Low-temperature structure and magnetic properties of the spinel LiMn₂O₄: A frustrated antiferromagnet and cathode material, Chem. Mater. 11, 1510 (1999).
  83. [83]J. B. Goodenough, Theory of the role of covalence in the perovskite-type manganites [La,M(II)]MnO₃, Physical Review 100, 564 (1955).
  84. [84]P. W. Anderson, New approach to the theory of superexchange interactions, Physical Review 115, 2 (1959).
  85. [85]J. B. Goodenough, An interpretation of the magnetic properties of the perovskite-type mixed crystals La₁₋ₓSrₓCoO₃₋ₓ, Journal of Physics and Chemistry of Solids 6, 287 (1958).
  86. [86]P. Patra, N. I. Naik, H. Bhatt, and S. D. Kaushik, Structural, infrared spectroscopy and magnetic properties of spinel ZnMn₂O₄, Physica B: Condensed Matter 10.1016/j.physb.2019.08.005 (2019).
  87. [87]R. Vidya, P. Ravindran, P. Vajeeston, H. Fjellvåg, and A. Kjekshus, Structural stability, electronic structure, and magnetic properties of mixed-valence acr3o8 phases (a = Na, k, rb), Phys. Rev. B 72, 014411 (2005).DOI
  88. [88]L. C. Chapon, P. G. Radaelli, G. R. Blake, S. Park, and S.-W. Cheong, Ferroelectricity induced by acentric spin-density waves in YMn2O5, Phys. Rev. Lett. 93, 177402 (2004).pubmed.ncbi.nlm.nih.gov/16606315
  89. [89]Y. Doi and Y. Hinatsu, Crystal structures and magnetic properties of ordered perovskites Sr2LnRuO6 (Ln = Eu–Lu), Journal of Physics: Condensed Matter 11, 4813 (1999).DOI
  90. [90]J. Gong, S. Dong, and J. Zhan, Understanding the multiferroicity in TmMn2O5 by a magnetically induced ferrielectric model, Sci. Rep. 6, 33529 (2016).

Paper details

vlnc:W_rxw8x7mh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents