Imagine a world where magnets work even at extreme temperatures, power better motors, or store data more efficiently. That is the goal of scientists using artificial intelligence (AI) to search for new magnetic materials. In a study published in Nature Communications, authors including Y. Zhang and colleagues created the NEMAD database with 67,573 records on magnetic materials. This database draws on scientific articles from Elsevier and the American Physical Society, with AI analyzing the texts and extracting details such as chemical composition, Curie temperature, or crystal structure.
The AI process begins with large language models (LLMs), such as GPT-4o. These models process articles in XML or PDF format, identify relevant sections, and create structured data. For example, the model divides text into 500-token chunks, finds similarities using the FAISS tool, and extracts information about materials such as Fe3PdN or VFeCoGe. This makes it possible to quickly compile vast amounts of data that would take years to collect manually. The NEMAD database contains elements such as iron (Fe), cobalt (Co), and nickel (Ni), and includes both ferromagnetic and antiferromagnetic compounds.
Training AI models for predictions
Using this data, scientists train machine learning models to predict material properties. They use Random Forest, XGBoost, and an Ensemble Neural Network. These models classify materials as ferromagnetic (FM), antiferromagnetic (AFM), or nonmagnetic (NM) with 90% accuracy. For example, XGBoost achieves an accuracy of 0.91 on test data, where key features such as average atomic mass or the proportion of elements with a high Curie temperature play a major role.
To predict the Curie temperature (the point at which a material loses its magnetism), the models work with balanced data to avoid bias toward low temperatures. XGBoost achieves a coefficient of determination R² of 0.87, a mean absolute error (MAE) of 56 K, and a root mean square error (RMSE) of 97 K. This means the model predicts temperatures with an error of around 56 kelvins. Similarly, for the Néel temperature (for AFM materials), R² is 0.83 and MAE is 38 K. The models take into account factors such as the proportion of iron or the average magnetic moment, which helps identify promising candidates.
25 new discoveries
The most interesting part is how AI searches external databases such as the Materials Project or databases of Heusler compounds to find new materials. The models identified 25 ferromagnetic candidates with a predicted Curie temperature above 500 K, such as GaFe2Co4Si with a prediction of around 1,000 K or Fe3Rh at 645 K. Of these, 7 were later verified in the literature, such as Fe2CuGa with an experimental temperature of 798 K. For antiferromagnetic materials, they found 13 candidates with a Néel temperature above 100 K, such as Sr2FeBrO3 at 284 K.
This approach accelerates discovery because AI filters thousands of compounds and proposes those with high potential. For example, the XGBoost model used balanced datasets in which it divided temperatures into bins and undersampled excess low-temperature data, improving accuracy by 4%. Scientists can thus focus experiments on promising materials such as VFe2BO5 or MnZn3(CrSe2)8 instead of testing randomly.
Thanks to AI, the discovery of magnetic materials is becoming faster and more efficient. The NEMAD database is available at www.nemad.org, where users can explore its records. The models not only make predictions but also reveal relationships, such as the importance of oxygen in antiferromagnets due to superexchange interactions. This approach can be extended to other fields, such as superconductors or photovoltaics, and helps address problems such as the shortage of rare earth elements in magnets.
