68,000 candidates, 4 new materials, all experimentally verified.

On July 3, Alibaba DAMO Academy, together with Renmin University of China and University of Chinese Academy of Sciences, released ElementsClaw, the industry's first AI Agent for superconductor discovery. This AI system predicted 68,000 potential superconducting materials, of which 4 new materials have been synthesized in the lab and confirmed to exhibit superconductivity. All data has been released openly for further exploration by the global research community. This isn't another story of AI publishing papers — it's a case of AI genuinely making scientific discoveries in the lab.

From “Guessing” to “Finding”: AI Redefines Materials Science

The history of superconductor discovery is full of accidents. In 1911 Dutch physicist Onnes discovered mercury's superconductivity through an experimental accident; the 1986 discovery of copper-based high-temperature superconductors won a Nobel Prize but was full of trial and error; in 2023 the Korean LK-99 room-temperature superconductor fiasco had the whole world watching a round of “mass alchemy.” The traditional materials science paradigm is: scientists propose hypotheses based on experience, then test them one by one in the lab — extremely inefficient, heavily dependent on intuition and luck.

ElementsClaw changes the rules of this game. It doesn't guess randomly; it systematically screens potential superconductor combinations through deep learning of periodic laws, crystal structures, and electron interactions. The 68,000 candidates aren't blind guesses — they're a “high-probability shortlist” AI generates from understanding physical chemistry laws, shrinking scientists' search space from astronomical numbers to experimentally verifiable ranges.

Most critically, the words “experimentally verified.” AI for Science has produced plenty of papers in recent years, but many remain at the “computational prediction” stage; few actually make it into the lab and get experimentally confirmed. DAMO's result has 4 new materials synthesized with measured superconducting signals, proving AI predictions aren't just academic exercises.

The Commercial Path for AI for Science

A separate QbitAI report the same day noted that the Physical AI concept Jensen Huang repeatedly emphasizes is being brought into life science labs by Chinese companies. AI for Science is moving from academic concept to industrial tool: drug discovery, materials design, protein folding, climate simulation, quantum computing control — all being accelerated by AI.

The market value of materials science is enormous. If superconductors can be realized at room temperature and ambient pressure, they would completely transform power transmission, MRI, quantum computing, and controlled nuclear fusion. Even “traditional” high-temperature superconductors have rigid demand in fusion tokamaks, MRI, and particle accelerators. AI accelerating materials discovery means these applications may commercialize sooner than expected.

DAMO's choice to fully open the data is also a smart strategy. Scientific discovery is inherently a public good; open data can attract global researchers to validate and iterate, accelerating the entire field. At the same time, Alibaba can build brand recognition in AI for Science through this achievement, attracting talent and collaboration opportunities.

AI for Science isn't a new story, but in 2026 it's moving from “demo” to “productivity.” DeepMind's AlphaFold transformed protein structure prediction; now systems like ElementsClaw are starting to transform materials discovery. When AI can genuinely make discoveries in the lab — not just draw tables in papers — the scientific research paradigm has truly changed.