NEMAD Database: Experimental Magnetic Resources

• NEMAD Database is an experiment-based magnetic materials repository that compiles chemical, structural, and magnetic property data from literature.
• It integrates LLM-assisted extraction and FAISS retrieval to achieve high precision (0.97) in extracting experimental magnetic labels.
• The database supports machine learning for magnetic state classification, transition temperature regression, and graph-neural-network applications after ICSD alignment.

NEMAD Database, or the Northeast Materials Database, is an experiment-based magnetic materials database developed to support data-driven discovery of magnetic compounds, especially compounds with high magnetic phase transition temperatures. It is presented as a large, structured, publicly accessible resource hosted at http://www.nemad.org, and its core contents include chemical composition, magnetic phase transition temperatures, structural details, magnetic properties, and provenance extracted from the scientific literature by LLM-assisted workflows. In the published literature, NEMAD serves both as a repository of experimentally sourced magnetic labels and as a substrate for machine-learning pipelines that classify magnetic state, predict Curie and Néel temperatures, and, after external structure alignment, enable graph-neural-network models for magnetic materials discovery (Itani et al., 2024, Schoener et al., 31 Jan 2026).

1. Scientific role and conceptual positioning

NEMAD was introduced to address a specific bottleneck in magnetic materials informatics: the lack of an accurate, comprehensive, and feature-rich database built from experimental reports rather than primarily from density functional theory calculations or small manually curated collections. The underlying motivation is that magnetic materials are central to permanent magnets, spintronics, medical technologies, data storage, and energy applications, while conventional discovery by intuition and experiment is slow and expensive. The database is therefore positioned as infrastructure for accelerated discovery of compounds with greater operating temperature ranges and optimized magnetic performance (Itani et al., 2024).

The database is explicitly distinguished from purely computational repositories such as the Materials Project, AFLOW, AFLOWLIB, and OQMD. The published rationale is that DFT is often unreliable for magnetic systems, especially strongly correlated and itinerant magnets, and that predicting Curie temperature from first principles typically requires additional model-Hamiltonian or mean-field approximations. NEMAD is also distinguished from smaller experimental resources such as MAGNDATA: the NEMAD paper characterizes MAGNDATA as containing about 2000 magnetic materials, with emphasis on lattice and magnetic structure rather than machine-learning-oriented breadth of features (Itani et al., 2024).

A common misconception is to treat NEMAD as a crystallographic archive. Later methodological work instead characterizes it as an experiment-based, literature-extracted database whose strength lies in experimentally grounded magnetic-property labels, while its original limitation is the absence of full atomic-coordinate crystal structures required by modern structure-aware graph models. In that sense, NEMAD occupies a hybrid position: it is richer in experimental magnetic labels than standard structure databases, but poorer in atomistic structural completeness than CIF-based repositories such as the ICSD (Schoener et al., 31 Jan 2026).

2. Literature extraction pipeline and curation workflow

NEMAD was constructed through an LLM-based literature extraction workflow centered on the authors’ GPTArticleExtractor method. The raw corpus began with about 40,000 DOIs of scientific articles related to magnetic materials, assembled by searching keywords such as ferromagnetic, antiferromagnetic, Curie, and Néel and by using web-scraping tools on journal websites. The paper states that the scholarly sources used for database construction were experimental articles published in Elsevier journals. Most articles were obtained in XML format, then processed with a custom XML parser that extracted full article content, including text and tables, and converted the material into plain text in CSV form with fields for DOI, title, abstract, and body text (Itani et al., 2024).

The initial GPTArticleExtractor pipeline segmented article text into 500-token chunks, embedded those chunks in a vector space, and used FAISS to retrieve the five most relevant segments for a given property. Prompt engineering and one-shot prompting were then used with GPT-3.5 and GPT-4 to produce structured JSON outputs. The target fields were chemical composition; magnetic phase transition temperatures (Curie, Néel, and Curie–Weiss); structural details (crystal structure, lattice structure, lattice parameters, space group); and magnetic properties (coercivity, magnetization, magnetic moment, remanence, susceptibility) (Itani et al., 2024).

The workflow was later revised with GPT-4o. Rather than relying on chunk retrieval alone, the revised approach supplied essentially the whole article except the introduction and references. The model output was required to be a list of dictionaries, each dictionary corresponding to a single material composition and its associated properties. This change was introduced because chunk retrieval could miss information in papers discussing multiple materials or variable stoichiometries. The revised full-article strategy is reported to have produced outputs that were more comprehensive, precise, and better organized (Itani et al., 2024).

Curation did not remain fully automated. After extraction, the database was manually validated. A key inclusion rule was imposed: if a chemical composition was extracted but lacked phase transition temperature information, it was excluded from NEMAD. Temperature values reported in different units were normalized to Kelvin. If a temperature was reported as a range, the range was replaced by the mean of the lower and upper bounds. Extraction quality was quantified by manual inspection of 200 randomly selected records, for which the authors report an overall precision of 0.97 (Itani et al., 2024).

3. Schema, coverage, and corpus composition

The published NEMAD schema contains 15 fields spanning composition, transition temperatures, structural descriptors, magnetic observables, provenance, and an experimental-status indicator. The database is therefore not merely a temperature table; it is structured as a multi-attribute magnetic materials resource (Itani et al., 2024).

Field	Type / unit	Function
Material_Chemical_Composition	string	Composition record
Curie	numeric, K	Curie temperature
Néel	numeric, K	Néel temperature
Curie Weiss	numeric, K	Curie–Weiss temperature
Crystal Structure	string	Structural descriptor
Lattice Structure	string	Structural descriptor
Lattice Parameter	string	Structural descriptor
Space Group	string	Structural descriptor
Coercivity	numeric, Oe	Magnetic property
Magnetization	numeric, A/M	Magnetic property
Magnetic Moment	numeric, $\mu_B$	Magnetic property
Remanence	numeric, A/M	Magnetic property
Susceptibility	numeric	Magnetic property
DOI	string	Literature provenance
Experimental	boolean Yes/No	Experimental-status field

The main NEMAD paper reports 26,706 magnetic materials spanning 84 different elements. Around 65% of records are ferromagnets, about 33% are antiferromagnets identified by materials with only a Néel temperature, and about 3% have both Curie and Néel temperatures. Approximately 40% of compounds are ternaries, with the remainder mainly quaternary, quinary, and binary compounds. Frequent elements include Fe, Co, Ni, Mn, La, Sr, Cu, B, Al, Cr, and Ce. The paper also highlights that the database contains a large number of magnetic compounds without rare-earth elements, and that roughly 22% of compounds have Curie temperatures above 600 K (Itani et al., 2024).

The published record also contains a numerical discrepancy. The NEMAD introduction paper reports 26,706 magnetic materials, and a later explainable-AI study repeats that figure for a publicly available NEMAD compiled using LLMs. By contrast, the ICSD-alignment paper cites the NEMAD work as containing 67,573 magnetic materials entries. This suggests that “materials” and “entries” may not be used identically across the literature, but the papers do not resolve the difference explicitly (Itani et al., 2024, Ajaib et al., 9 Aug 2025, Schoener et al., 31 Jan 2026).

4. Baseline machine learning and candidate screening

The original NEMAD study couples the database to composition-based machine learning for both magnetic-state classification and transition-temperature regression. The core representation is an 84-dimensional elemental proportion vector over all elements present in the database. If $n_i$ is the number of atoms of element $i$ and $N$ is the total number of atoms in the formula, the elemental proportion is defined as

$p_i = \frac{n_i}{N}.$

Additional handcrafted compositional features include average atomic number, average atomic weight, average electronegativity, average atomic magnetic moment, average group, average period, the $L2$ stoichiometry norm, a configurational entropy-like quantity, the total proportion of high-Curie magnetic elements (Fe, Co, Ni), and the total rare-earth proportion. The printed formulas include

$\bar{Z} = \sum_{i=1}^{84} p_i Z_i,$

$L2 = \sqrt{\sum_{i=1}^{84} p_i^2},$

and

$\bar{S} = -\sum_{i=1}^{84} p_i \log p_i.$

For structure-informed models, encoded crystal system and space group were added as two more features (Itani et al., 2024).

For magnetic-state classification, the authors supplemented NEMAD with 11,389 stable nonmagnetic materials from the Materials Project. After combining and deduplicating, they report 28,168 unique entries. A Random Forest classifier trained with a 60%/20%/20% train/validation/test split and hyperparameters selected by five-fold cross-validation achieved 0.90 accuracy on both validation and test sets. On the testing set, classwise results were: FM precision 0.90, recall 0.90, F1 0.90; AFM precision 0.82, recall 0.76, F1 0.79; and NM precision 0.93, recall 0.96, F1 0.94. The AFM class was consistently weaker, which the paper attributes to fewer AFM examples (Itani et al., 2024).

For transition-temperature regression, the paper evaluates Random Forest, XGBoost, and Ensemble Neural Network models. The ferromagnetic dataset had 11,923 unique compounds, and because about 51% of compounds were below 300 K while about 22% were above 600 K, the authors created a balanced ferromagnetic dataset of 8,249 compounds by randomly removing low-temperature records below 300 K. The antiferromagnetic dataset had 5,389 unique compounds. The best Curie-temperature result was XGBoost on the balanced dataset, with $R^2 = 0.86$, MAE = 62 K, and RMSE = 105 K. The best Néel-temperature result was also XGBoost, with $n_i$0, MAE = 32 K, and RMSE = 67 K. The paper also reports a structure-informed Curie model with $n_i$1, MAE = 52 K, and RMSE = 89 K (Itani et al., 2024).

Task	Best reported result	Notes
FM/AFM/NM classification	Accuracy 0.90	Random Forest
Curie prediction	$n_i$2, MAE = 62 K	XGBoost, balanced composition-only set
Curie prediction with structure	$n_i$3, MAE = 52 K	XGBoost, composition + structure
Néel prediction	$n_i$4, MAE = 32 K	XGBoost

NEMAD was also used for screening unseen compounds from the Materials Project. After removing compounds already present in NEMAD and retaining only compounds for which the authors’ magnetic-state classifier agreed with the Materials Project label, the screening set contained 602 ferromagnetic and 237 antiferromagnetic previously unseen compounds. Using thresholds of predicted $n_i$5 K for ferromagnets and predicted $n_i$6 K for antiferromagnets, the study identified 62 ferromagnetic candidates and 19 antiferromagnetic candidates. Representative FM candidates include FeCo$n_i$7Ge, AlFeCo$n_i$8, and Ga$n_i$9Fe$i$0Co$i$1Si; representative AFM candidates include TiFeO$i$2, Sr$i$3FeBrO$i$4, and La$i$5Mn$i$6Se$i$7O$i$8 (Itani et al., 2024).

5. Downstream methodological developments

Subsequent work uses NEMAD not only as a fixed benchmark but as a platform for methodological extensions. In the explainable-AI study on Curie-temperature prediction, NEMAD is described as a publicly available database compiled using LLMs and containing 26,706 magnetic materials with chemical composition, phase transition temperatures, structural characteristics, and related magnetic properties. That study pre-processed NEMAD-derived records to remove ambiguous entries, duplicates, and entries with missing or non-numeric temperatures, standardized chemical formulas using pymatgen, and then applied a balancing step by randomly removing around 4% of samples with Curie temperatures less than 300 K. Its best model, Extra Trees, reached cross-validated performance of MAE = 54 ± 2.5 K, RMSE = 105 ± 4.0 K, and $i$9 on the balanced dataset, while external validation on DS1 yielded approximately $N$0 and MAE $N$1 K (Ajaib et al., 9 Aug 2025).

That explainable-AI work also exposes NEMAD’s internal heterogeneity. Using k-means with $N$2, it identified chemically distinct groups including metallic intermetallics rich in Fe, Co, Si, and B; oxygen-rich complex oxides with O, Mn, La, and Sr; Mn/Fe/Ga/Ni-enriched Heusler-like alloys; and a cluster dominated by C, Fe, and Co. Predictive performance degraded substantially on some chemically specific subspaces, with Cluster 1 at $N$3 and Cluster 3 at $N$4. The paper attributes these failures to effects such as superexchange, strong electron correlation, Jahn–Teller effects, complex bonding environments, and non-collinear magnetic structures, all of which are poorly captured by composition-only descriptors. SHAP analysis ranked features such as Mean Magnetic Moment, Average Deviation of Number of Unfilled f-Orbitals, Mean Ground State Volume/Atom, Fraction of Fe, Fraction of Mn, and average deviation of magnetic moment among the most important predictors (Ajaib et al., 9 Aug 2025).

A second line of development addresses the structural incompleteness of NEMAD. The alignment paper argues that NEMAD originally lacks full atomic-coordinate crystal structures, preventing direct use of models such as CGCNN. To bridge this gap, it aligns NEMAD with ICSD crystallographic information files by normalizing chemical formulas with PyMatGen, matching reduced formulas, and then optionally checking agreement of space-group International Tables numbers. This yields two aligned subsets: Database 1 by composition-only matching, with 11,292 entries with Néel temperatures and 8,213 entries with Curie temperatures; and Database 2 by composition plus space group, with 5,147 entries with Néel temperatures and 3,821 entries with Curie temperatures. Structural ambiguity is quantified by a noise metric based on the spread of metric tensors of Niggli-reduced cells, and the reported noise drops from $N$5 for composition-only alignment to $N$6 for composition-plus-space-group alignment (Schoener et al., 31 Jan 2026).

The aligned datasets enable CGCNN models with an 80/10/10 train/validation/test split. Relative to the original NEMAD-based models, performance improves from Curie MAE = 56 K, Néel MAE = 38 K, and CCR = 0.90 to Curie MAE = 37.3 K, Néel MAE = 22.6 K, and CCR = 0.95 under the stricter alignment. A transfer-learning variant further reduces Néel MAE to 22.0 K. The methodological significance is that NEMAD becomes, after alignment, a property-labeled crystal-graph dataset rather than only a composition-and-metadata table (Schoener et al., 31 Jan 2026).

6. Limitations, unresolved issues, and access

The literature states several limitations explicitly. First, NEMAD’s extraction corpus was mainly Elsevier experimental articles, so coverage is restricted by publisher scope. The original paper notes that many antiferromagnetic oxides are published outside Elsevier, including in APS journals, and presents this as one reason for the database’s imbalance toward ferromagnets. Second, extraction errors still occur, especially when a single paper contains many compounds or variable stoichiometries. Third, the transition-temperature datasets are imbalanced, especially in the high-temperature regime, which leads models trained on the original distribution to underestimate Curie temperatures above 500 K and to deviate more strongly in sparse high-temperature regions (Itani et al., 2024).

A further limitation is structural incompleteness. In its original form, NEMAD may include chemical composition and partial structural descriptors such as crystal system, lattice type, or space group, but not the complete structural information needed to build atomistic graphs. Later alignment work shows that linking NEMAD to ICSD reduces this problem but does not eliminate ambiguity: multiple ICSD CIFs can share the same composition and even the same space group, a single NEMAD entry can map to multiple candidate structures, and when multiple NEMAD rows match a CIF, one row is randomly chosen. The alignment workflow is therefore powerful but not exact (Schoener et al., 31 Jan 2026).

The literature also leaves some interpretive issues open. Most notably, the database size is reported inconsistently across papers as 26,706 magnetic materials and 67,573 magnetic materials entries. The papers do not provide a formal reconciliation. A plausible implication is that the database evolved across releases or that entry counts and material counts are not equivalent, but this remains an inference rather than a documented resolution (Itani et al., 2024, Schoener et al., 31 Jan 2026).

In terms of access, the NEMAD paper emphasizes a user-friendly website at http://www.nemad.org through which users can explore and retrieve detailed information about each entry. The papers do not describe a programmable public API for NEMAD itself. What is documented is a public web presence, literature-derived provenance through DOI, and repeated use of NEMAD as a public database in downstream methodological studies (Itani et al., 2024, Ajaib et al., 9 Aug 2025).

Taken together, these publications define NEMAD as a magnetic materials data infrastructure rather than merely a static table: an experiment-based, LLM-assisted literature database that supports composition-based prediction, explainable modeling, and, after external structure alignment, graph-neural-network learning. Its chief value lies in experimentally sourced magnetic transition-temperature labels at scale; its chief constraints are publisher-bounded extraction, residual curation noise, class imbalance, incomplete atomistic structure, and unresolved versioning or counting ambiguities in the published descriptions.