In the field of physics, exploring superconductors has been described as a century-long quest for the "Holy Grail of Physics." Recently, this traditional area of research welcomed a powerful "AI teammate" — an AI agent called "ElementsClaw" (Elemental Crab), specifically designed to discover superconductive materials. In just 28 GPU hours, the system conducted a comprehensive scan of 2.4 million stable crystals and identified 68,000 potential superconductors, with a research efficiency far surpassing that of human exploration over centuries.

For a long time, the search for superconductors has relied on the "cookbook-style" approach, where different element combinations are tested through trial and error. This method not only has a very low success rate but also often depends on accidental discoveries. To change this situation, the research team developed a "convergent expertise" architecture, endowing ElementsClaw with unique research capabilities: its built-in 1 billion parameter geometric deep graph neural network can accurately interpret three-dimensional crystal structures; combined with a large language model, the AI agent can autonomously read literature, access data, and assist in decision-making.

image.png

This AI system is not only a powerful tool for "finding a needle in a haystack," but it also has self-evolution capabilities. During the experimental validation phase, researchers successfully synthesized four entirely new superconductors previously unknown to humans, with each material's discovery showcasing different logical paths taken by the AI. From re-evaluating forgotten database structures, to correcting past computational errors, to generalizing from structural motifs, ElementsClaw demonstrated a leap from "auxiliary judgment" to "active design" in the field of materials science.

Although the critical temperatures of these newly discovered materials have not yet reached room temperature standards, their core value lies in successfully validating the application path of AI agents in this field. Compared to the natural superconductive material hit rate of about 3%, ElementsClaw's recommendation accuracy has improved by an order of magnitude.