A large, rich, and experimentally based database allows models to quickly search for patterns and suggest new candidates. The result is an online database built with large language models that read and extract data from a vast body of scientific literature.
Strong magnets are found almost everywhere. In electric vehicle motors, wind turbine generators, speakers, sensors, and industrial systems. The problem is that many advanced magnets rely on rare earth elements. This creates a dependency on limited supply chains, adding uncertainty to prices and availability. Therefore, “alternative” magnetic materials, which provide good performance without a high dependence on rare earth elements, have become a major research target.
The new study suggests a different approach to finding such materials. Instead of starting with a small number of candidates and testing them in the lab, the researchers first built the infrastructure: a large, rich, experimentally-based database that allows the models to search for patterns and quickly suggest new candidates. The result is an online database built using large-scale lip models that read and extract data from a vast body of scientific literature.
A database based on experiments, not just calculations
The database includes 67,573 records of magnetic materials. Each record is organized around a clear material identity, i.e. chemical composition, structural information, and magnetic properties. Each material also contains many fields designed to facilitate machine learning. These include details about transition temperatures, such as the Curie temperature or Néel temperature, as well as other parameters that are important for understanding magnetic behavior.
The researchers emphasize that the value of the database is that it is based on experimental data reported in the papers, and not just on computational predictions. In the field of materials, this is a critical difference. Sometimes calculations identify a “promising” magnet, but when it gets to the laboratory, problems with stability, impurities, or different crystallization phases are discovered. A database that relies on experimental results provides a better starting point.
To build it, the researchers collected a large volume of experimental articles, broke them down into text segments, and used verbal models to extract specific fields from them. This is a complex process, because the information does not always appear in the same format. Sometimes the data is in tables, sometimes in continuous text, and sometimes you need to identify the relationship between a particular substance and a measurement made on it under certain conditions. This is where the power of models comes in: they can perform systematic extraction on a large scale, which is not possible by manual work.
What did they do with the database, and what does it mean for the industry?
Once the database was established, the researchers trained models on it that classify materials and evaluate properties, then ran the models on external databases to search for new candidates. They reported a group of materials that the models identified as having the potential to be magnetic even at high temperatures. With an emphasis on temperatures above 500 Kelvin, that is, above about 227 degrees Celsius.
Here it is important to make a reservation. A “good” magnet for industry is not just a material with a high transition temperature. It also needs properties such as coercivity, magnetization, thermal and mechanical stability, and the ability to produce the material industrially and at a reasonable cost. The database does not replace experimentation. It is supposed to shorten the path to experimentation, and prevent wasting time on weak directions.
The practical importance is in creating a “fast track.” When there is a large experimental database, it is possible to run models on it, obtain lists of candidates, and direct them to laboratories that are capable of synthesizing and measuring. In this way, basic research can more quickly connect to the needs of industries such as electric vehicles and renewable energy. If the approach expands to other fields, it could change the way functional materials are searched for in general, not just magnetic ones.
To the scientific article (Nature Communications)
More of the topic in Hayadan:
