Machine-accelerated extraction of experimental superconducting parameters for physical insight and prediction across thousands of materials
This paper introduces AutoSuperCon, a large-scale database of 20,354 experimentally validated superconducting parameters extracted via large language models from over 47,000 arXiv preprints, which enables new physical insights and accurate machine learning predictions for thousands of superconductors.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
Superconductors are materials that can carry electricity without any resistance, meaning no energy is lost as heat. This property makes them the holy grail for technologies ranging from powerful magnets in medical scanners to the lossless power grids of the future. For a material to be useful in these applications, however, knowing just the temperature at which it becomes superconducting is not enough. Engineers also need to know how much magnetic field the material can withstand before it stops working, and how much electrical current it can carry. For decades, finding new superconductors has been a game of chance rather than a precise science, with major discoveries like copper-based and iron-based superconductors appearing without any prior theoretical prediction. While computers have recently been used to guess which chemical recipes might work, they have been limited by a lack of reliable data on the other crucial properties beyond that initial temperature.
A team of researchers has now built a massive new digital library to solve this problem, creating a resource that could finally allow computers to learn the full story of superconductivity. They developed an automated system that read through more than 47,000 scientific papers published on a public archive, using advanced language software to pull out specific experimental numbers. The result is a database called AutoSuperCon, which contains over 20,000 entries for nearly 7,000 different superconducting materials. This collection includes not just the temperature at which a material becomes superconducting, but also eight other key measurements, such as the maximum magnetic field it can survive, the depth to which magnetic fields can penetrate it, and the size of the energy gap that allows electrons to pair up. To ensure the data was accurate, the team cross-checked their findings against existing records, finding that their automated extraction matched human-curated data with high precision. They also included historical measurements from a 1976 survey that had previously been locked away in a scanned document that standard computers could not read, digitizing it with the help of modern vision software.
With this vast new dataset in hand, the researchers were able to test long-standing theories about how superconductors behave. They found that while older theories predicted a strict limit on how strong a magnetic field a superconductor could tolerate, nearly 30 percent of the materials in their database actually exceed this limit. This suggests that many modern superconductors operate in a regime that standard theories do not fully explain. They also discovered that the energy gap in these materials is often much larger than expected, particularly in the copper-based and iron-based families, indicating that the forces holding the electron pairs together are much stronger than in conventional materials. Furthermore, the team identified a clear upper ceiling for the operating temperature of these materials, which appears to be directly linked to the density of the superconducting fluid within the material. This relationship holds true across thousands of different compounds, providing a new quantitative rule for what is physically possible.
The most immediate application of this work is in the ability to predict the properties of new materials before they are even made. The researchers trained computer models to guess the critical temperature and the maximum magnetic field a material can handle based solely on its chemical ingredients. These models were highly accurate for materials within known families, correctly predicting the critical temperature for thousands of held-out examples. However, the study also revealed a hard limit to this approach: the models failed completely when asked to predict properties for entirely new families of materials that were not part of their training. This means that while the database is a powerful tool for screening candidates within known chemical territories, it cannot yet predict the properties of a completely unknown class of superconductors. The work demonstrates that artificial intelligence can now extract and organize the hidden knowledge buried in decades of scientific literature, turning scattered experimental notes into a structured foundation for discovering the next generation of superconducting materials.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.