An Alibaba research agent just helped scientists find four new superconductors

Alibaba's DAMO Academy, Renmin University and UCAS say their new AI agent has done something the materials community has chased for years: it designed a batch of superconductors end-to-end and had four of them confirmed in a real lab.

The system, called ElementsClaw, is built on a 1-billion-parameter atomic foundation model called Elements that is wrapped inside an LLM agent. The agent reads the literature, scores whether a candidate can actually be synthesized, and proposes an experiment plan. Then humans run the experiment. On July 3, the team released both the methodology paper on arXiv (2604.23758) and the full 2.4-million-crystal screening database on the DAMO AI-for-Science portal, with the announcement carried the same week by IT之家 and 第一财经.

The headline number is the loop time. The agent took 28 GPU-hours to triage roughly 2.4 million crystal structures down to about 68,000 candidates that looked both physically plausible and worth synthesizing. From there, the team picked the most promising and tested them. Four synthesized successfully: Hf21Re25, Zr4VRe7, HfZrRe4, and Zr3ScRe8, all hafnium- and zirconium-rich rhenium compounds. The best of them superconducts at about 6.5 Kelvin. That is cold, but it is a real, measured transition temperature, not a model score.

The mechanism underneath is what makes the release worth more than a novelty item. The Elements model itself posts near-perfect scores on standard superconductivity classification tests (an AUC of 0.996 in the team's reported benchmark) and predicts critical temperatures with a mean error under one Kelvin. The agent layer on top is the new part. It automates the literature review and the synthesis-feasibility check, then hands a chemist a shortlist instead of a million-row spreadsheet. As DAMO science-AI lead Rong Yu put it in the IT之家 coverage, this is the first batch of superconductors that an AI agent both discovered and had verified in a lab.

Read it that way and the release is a marker for where AI-for-science has moved. The bottleneck used to be running the prediction. Now the model can rank the periodic table faster than a grad student can read abstracts, so the scarce resource is which question to give the agent. The team's open materials database is also a bet on that shift: releasing 2.4 million scored structures gives other groups a head start on the same triage problem for solid-state battery electrolytes, multi-phase catalysts, and thermoelectrics.

The arXiv listing was confirmed for title and authors, but the abstract body was truncated in the available materials, so the paper itself is the strongest primary anchor and should be checked before any quantitative claim is republished. The 2,000-material baseline that the team uses to frame "screening 2.4 million" is a company claim restated in 第一财经, not an independent benchmark. "First AI-agent-discovered and lab-verified batch of superconductors" is DAMO's framing of its own result, and should be read as that.

What to watch next: whether the same ElementsClaw recipe produces a lab-verified hit in a hotter field, like a room-temperature-superconductor candidate or a solid-state electrolyte, and whether outside groups can replicate the 68,000-candidate shortlist on the open DAMO database without re-tuning the agent. The next paper, not the next press release, is the test.