The Northeast Materials Database for Magnetic Materials

This study introduces the Northeast Materials Database (NEMAD), a comprehensive resource of 67,573 magnetic materials entries generated via Large Language Models, and demonstrates its utility in training machine learning models that achieve high accuracy in classifying magnetic types and predicting transition temperatures to accelerate the discovery of high-performance magnetic materials.

The Gist

Imagine you are a chef trying to invent the perfect new recipe for a dish that needs to stay hot for a very long time (like a magnet that works in a scorching engine). For centuries, chefs (scientists) have been finding these recipes by taste-testing one by one. It's slow, expensive, and they often run out of rare, expensive ingredients (like rare-earth elements).

Recently, a new team of chefs decided to try a different approach. Instead of tasting every dish, they built a massive digital cookbook using a super-smart AI assistant. Here is how they did it, explained simply:

1. The Problem: The Library is Too Big

For a long time, scientists knew that magnetic materials (the stuff inside your hard drive or a wind turbine) were great, but the best ones were rare and broke down when they got too hot. To find new, better ones, they needed to read millions of scientific papers. But these papers are written in complex language, scattered across different books, and full of messy tables. Reading them all by hand would take a human lifetime.

2. The Solution: The AI Librarian (LLMs)

The researchers at the University of New Hampshire hired a "super-librarian" powered by Large Language Models (LLMs)—the same kind of AI that powers chatbots.

Think of this AI as a robot that can read 100,000 scientific papers in seconds. But instead of just reading them, the robot was taught to extract the specific ingredients from the text. It looked for:

• The Recipe: What elements are in the material? (e.g., Iron, Cobalt, Oxygen).
• The Structure: How are the atoms arranged? (Like the shape of the kitchen).
• The Performance: How hot can it get before it stops working? (The "Curie Temperature").

The AI didn't just read; it organized this chaos into a neat, searchable spreadsheet called NEMAD (The Northeast Materials Database). It pulled data from 67,573 different magnetic materials, turning messy PDFs and scanned old handbooks into clean, digital data.

3. The Training: Teaching the AI to Predict

Once the database was built, the team taught a new set of AI models (Machine Learning) to look at a "recipe" and guess the result.

• The Sorter (Classification): They taught the AI to look at a list of ingredients and instantly say: "Is this a magnet? If so, is it a Ferromagnet (sticks to your fridge) or an Antiferromagnet (a more complex, invisible magnet)?"

  • Result: The AI got this right 90% of the time. It's like a master chef who can smell a dish and know exactly what spices are in it without tasting it.

• The Predictor (Regression): They taught the AI to guess the "breaking point" temperature. "If I mix Iron, Cobalt, and this specific crystal shape, how hot will it get before it loses its magic?"

  • Result: The AI's guesses were very close to reality, with an error margin of only about 50 degrees (which is pretty amazing for guessing the future of atoms).

4. The Discovery: Finding the "Golden Nuggets"

Now that the AI was smart, they used it to scan a massive list of potential materials that no one had tested yet (from a database called the "Materials Project").

The AI acted like a metal detector in a giant field. It scanned thousands of theoretical recipes and shouted: "Stop! These 25 materials look like they can handle temperatures over 500 Kelvin (440°F)!"

• The Proof: They checked the literature and found that for some of these "AI-predicted" materials, humans had already tested them and confirmed the AI was right.
• The Future: The other 25 are brand new candidates. They are the "hidden gems" waiting for real-world scientists to go into the lab, mix the chemicals, and see if the AI was right.

Why This Matters

This paper is a game-changer because it changes the speed of discovery.

• Before: Scientists spent years guessing and testing one material at a time.
• Now: They have a digital map of the entire landscape of magnetic materials. They can use AI to instantly spot the best paths to new, high-performance magnets that don't rely on rare, expensive elements.

In short: They built a super-smart robot librarian to read the world's scientific books, organized the data into a giant database, and then used that brainpower to predict where the next generation of super-magnets is hiding. It's like using a GPS to find a treasure chest instead of digging a hole in the backyard.

Technical Summary

1. Problem Statement

The discovery of high-performance magnetic materials with wide operating temperature ranges is critical for applications in renewable energy, data storage, and quantum computing. However, current research faces three primary bottlenecks:

• Data Scarcity and Fragmentation: Existing magnetic material databases (e.g., MAGDATA) are small (approx. 2,000 entries), lack comprehensive features (missing structural details and specific magnetic properties), and rely on manual curation.
• Limitations of Computational Methods: While Density Functional Theory (DFT) is widely used, it often fails to accurately predict magnetic properties for strongly correlated electron systems (like itinerant magnets) and is computationally expensive for large unit cells.
• Data Extraction Challenges: Scientific literature contains vast amounts of experimental data, but it is unstructured (PDFs, tables, scanned images). Traditional Natural Language Processing (NLP) tools struggle to extract complex, multi-variable data (e.g., variable stoichiometry like $Fe_xCo_{1-x}$) and structural details from these formats accurately.

2. Methodology

The authors developed a comprehensive pipeline to construct the Northeast Materials Database (NEMAD) and train machine learning (ML) models for material classification and property prediction.

A. Database Construction (NEMAD)

• Data Sources: The team compiled 100,000 Digital Object Identifiers (DOIs) from Elsevier and American Physical Society (APS) journals.
• Hybrid Extraction Workflow: To overcome the limitations of previous tools, they implemented a multi-pathway extraction strategy using Large Language Models (LLMs), specifically GPT-4o and Google Gemini 2.0 Flash:
  • XML/API: Parsed directly from journal APIs using custom scripts.
  • Standard PDFs: Converted to Markdown using a PDF parser to preserve layout (tables, headers) before LLM processing.
  • Scanned PDFs/Handbooks: Used OCR capabilities of Gemini 2.0 Flash to convert images to structured Markdown.
• Prompt Engineering & Optimization:
  • Utilized a two-stage retrieval system (using FAISS for vector similarity) to identify the most relevant text segments within long articles, mitigating token limits.
  • Employed one-shot prompting with illustrative examples to ensure consistent JSON output formats.
  • Addressed variable stoichiometry by refining prompts to explicitly capture variable values (e.g., $x$ in $Fe_xCo_{1-x}$).
• Data Validation: A subset of 5,015 records was independently validated using Google Gemini 2.5, achieving a median accuracy of 94%.
• Scope: The final database contains 67,573 entries, including chemical composition, structural details (crystal system, lattice parameters, space group), and magnetic properties (Curie/Néel temperatures, coercivity, magnetization, etc.).

B. Feature Engineering

• Chemical Composition: Converted chemical formulas into an 84-dimensional elemental proportion vector. Derived features included average atomic number, weight, electronegativity, magnetic moment, L2 stoichiometry norm, and entropy.
• Structural Features: Applied one-hot encoding for crystal systems (7 types) and label encoding (based on target variable averages) for space groups to handle high dimensionality.

C. Machine Learning Models

• Classification: Trained Random Forest (RF) and XGBoost classifiers to categorize materials as Ferromagnetic (FM), Antiferromagnetic (AFM), or Non-Magnetic (NM). The dataset was augmented with 11,389 non-magnetic compounds from the Materials Project.
• Regression: Trained RF, XGBoost, and Ensemble Neural Networks (ENN) to predict Curie ($T_C$) and Néel ($T_N$) temperatures.
• Handling Data Imbalance: Recognizing the skew toward low-temperature data, the authors employed stratified undersampling to create balanced datasets, ensuring better model performance across the full temperature spectrum.
• Ensemble Strategy: Trained ensembles of 30 models on different balanced subsets to generate robust predictions and estimate uncertainty via standard deviation.

3. Key Results

Database Statistics

• Size: 67,573 unique magnetic material entries.
• Composition: ~68% FM, ~30% AFM, ~2% mixed.
• Features: Includes 84 unique elements, with a significant portion of rare-earth-free compounds.
• Temperature Distribution: ~22% of compounds have $T_C > 600$ K.

Model Performance

• Classification:

  • Accuracy: Achieved 90% accuracy on the test set for distinguishing FM, AFM, and NM.
  • Robustness: Both RF and XGBoost showed consistent precision, recall, and F1-scores, with XGBoost slightly outperforming RF (0.91 vs 0.90 on test sets).
  • Feature Importance: Key predictors included average atomic weight, magnetic moment, proportion of high-$T_C$ elements (Fe, Co, Ni), and electronegativity.

• Regression (Curie Temperature Prediction):

  • Best Model: XGBoost on the balanced dataset.
  • Metrics: $R^2 = 0.87$, Mean Absolute Error (MAE) = 56 K, RMSE = 97 K.
  • Improvement: The balanced dataset improved $R^2$ by 4% compared to the original skewed dataset, specifically enhancing high-temperature prediction.

• Regression (Néel Temperature Prediction):

  • Best Model: XGBoost.
  • Metrics: $R^2 = 0.83$, MAE = 38 K, RMSE = 72 K.
  • Feature Insight: The proportion of Iron (Fe) and Oxygen (O) were identified as critical features for AFM behavior, consistent with superexchange interaction physics.

Material Discovery (Screening)

• The models were applied to screen the Materials Project and a set of Heusler compounds.
• Identified Candidates:
  • 25 Ferromagnetic candidates with predicted $T_C > 500$ K.
  • 13 Antiferromagnetic candidates with predicted $T_N > 100$ K.
• Validation: 7 of the identified candidates were cross-referenced with existing literature and confirmed to have experimental ordering temperatures, validating the model's predictive power.

4. Significance and Impact

• Scalable Data Curation: Demonstrates that LLMs can effectively automate the extraction of complex, structured scientific data from diverse document formats (XML, PDF, scanned images), overcoming the bottleneck of manual curation.
• High-Quality Resource: NEMAD is the largest and most feature-rich experimental magnetic materials database to date, filling a critical gap for data-driven materials science.
• Accelerated Discovery: The combination of a comprehensive database and robust ML models enables the rapid screening of vast compositional spaces (e.g., Materials Project) to identify high-temperature magnetic materials without rare-earth elements.
• Generalizability: The workflow (GPTArticleExtractor) is adaptable to other material science domains (superconductors, thermoelectrics, etc.), offering a blueprint for transforming how scientific knowledge is aggregated and utilized.

5. Availability

• Database: Accessible at www.nemad.org.
• Code & Data: Source code and processed datasets are available on GitHub and Zenodo.