Generating is not discovering: a pre-registered physics judge for AI-proposed superconductors, calibrated on six known superconductors and one negative control

The search for new superconductors—materials that conduct electricity with zero resistance—has entered a new era where artificial intelligence is asked to imagine them. For decades, scientists have known that for a material to become a superconductor, its electrons must pair up in a specific way, and the material must be strong enough to be drawn into a wire for practical use. Recently, powerful computer models have begun generating millions of new crystal structures, claiming they are stable and promising. However, a critical question remains unanswered: do these computer-generated ideas actually possess the physical properties needed to work, or are they just mathematically stable shapes that fail when tested against the laws of physics? The field has lacked a standard way to check if a proposed material can truly pair its electrons and survive the journey from a lab bench to a real-world application.

Reinaldo Inácio, a researcher based in São Paulo, has built a rigorous testing system to answer this question. He calls it a "physics judge." Unlike previous methods that simply check if a structure holds together, this judge simulates the complex behavior of electrons to see if they will pair up and form a superconducting state. The system takes a crystal structure as input and runs it through a series of seven strict stages. First, it calculates the electronic structure of the material. Then, it builds a simplified model to track how electrons move. Next, it calculates how the material responds to magnetic fluctuations, which are the forces that often cause electrons to pair up in high-temperature superconductors. Finally, it checks if the material can be made into a wire, looking at factors like how the electrons behave at the boundaries between crystals and how the material handles stress. Every step is governed by pre-set rules, and the entire process is designed to be transparent and reproducible, with every calculation recorded so that anyone can verify the result.

To prove that this judge works, the researcher first tested it on six known superconductors, including materials like niobium and a copper-based compound called YBa2Cu3O7. Before running the calculations, he wrote down exactly what the judge should find for each material, creating a set of "answer keys." The system successfully identified the correct pairing behavior for all six known materials, confirming that it could recognize the physics of real superconductors. The judge also correctly identified a negative control: a material that looks very similar to a known superconductor but does not actually superconduct. When the judge analyzed this non-superconducting material, it correctly determined that the electrons would not pair up, even though the material's structure was nearly identical to a working one. This showed that the judge is sensitive enough to distinguish between materials that work and those that do not, based on subtle differences in their electronic interactions.

The researcher then used this judge to audit the output of current artificial intelligence generators. He examined thousands of structures proposed by AI systems and ran them through the physics judge. The results were stark. When the judge analyzed 1,248 new structures generated by an AI system called MatterGen, it found zero candidates that were both new and capable of carrying the specific electron pairing needed for a wire. Similarly, when the judge scanned over 47,000 compounds from a public database, it found that while many had the right geometric shape, they lacked the necessary electronic strength to become superconductors. The study concluded that the current generation of AI tools is excellent at finding stable shapes, but it is not yet discovering new physics. The AI is essentially re-finding materials that scientists already know about, or proposing structures that look promising on paper but fail the physical test of electron pairing.

The paper argues that the bottleneck in discovering new superconductors is no longer the ability to generate ideas, but the ability to judge them. The cost of running this entire physics judge on a new material is surprisingly low, estimated at under 200 US dollars in cloud computing costs for the entire study. The researcher proposes that the scientific community adopt this judge as a public benchmark. Before any new AI-generated material is celebrated as a discovery, it should be run through this test to see if it truly has the physics to do the job. Until such a test is applied, the thousands of "stable" structures proposed by AI remain unproven ideas. The study suggests that true discovery will only happen when the generation of new ideas is matched by a rigorous, physics-based filter that can say "no" to materials that cannot work, ensuring that future efforts focus only on candidates with a genuine chance of becoming the next generation of superconducting wires.