News · 2026-07-04

An AI Agent Screened 2.4 Million Crystals and Found Four New Superconductors

An AI agent has screened 2.4 million candidate crystal structures and identified four new materials that a laboratory has since synthesized and confirmed are genuine superconductors. The system, called ElementsClaw, comes from Alibaba's DAMO Academy working with Zhejiang University and Westlake University, and the work has been accepted to ICML 2026. One of the four new materials was invented from scratch by the model rather than adapted from something already known.

Key facts

• ElementsClaw screened 2.4 million candidate crystal structures in about 28 GPU-hours and flagged roughly 68,000 high-confidence candidates.
• Four of its new candidates were synthesized and measured in a lab; all four superconduct, with critical temperatures of about 6.5 K, 5.9 K, 3.5 K, and 2.5 K.
• It also rediscovered 66 known superconductors missing from its reference database, a built-in check that its screening actually works.
• Published by Alibaba DAMO Academy, Zhejiang University, and Westlake University; accepted to ICML 2026 (arXiv 2604.23758).

A superconductor is a material that carries electricity with zero resistance, meaning current can flow through it forever without losing energy as heat. Only about 2,000 superconductors are known in total, discovered over more than a century of trial and error, so finding four new, lab-confirmed ones in one project is a meaningful addition to that short list.

For decades, finding a new superconductor has meant a slow loop of guessing a chemical formula, growing a crystal, cooling it down, and measuring whether current flows without resistance - a process that can take a researcher months per candidate. ElementsClaw compresses the guessing part of that loop. It pairs a language model, which does the semantic reasoning and planning (deciding which regions of chemical space look promising, based on patterns in how known superconductors are built), with a much smaller "Large Atomic Model," a 1-billion-parameter system that does the fast numerical physics of predicting how a candidate crystal will behave. The two run in a loop: the language model proposes and narrows, the atomic model checks the physics cheaply, and results feed back into the next round of proposals.

The everyday version of this is like a chef and a taste-tester working together at scale. Instead of one person alone slowly baking and tasting ten thousand cookie recipes, one much faster taster (the atomic model) rapidly screens which recipes are even worth baking, while the chef (the language model) uses judgment about flavor combinations to decide where to look next. That division of labor is what let the system get through 2.4 million candidates in roughly 28 GPU-hours - work that would take a human research group years of lab time to attempt even a tiny fraction of.

The team validated the approach in two ways. First, they checked whether the system could find superconductors that were known to exist but had been deliberately left out of its training data - it rediscovered 66 of them, evidence that its screening genuinely tracks real physics rather than memorized answers. Second, and more convincingly, they took four of the system's brand-new candidates - materials with no prior record in the literature - and had a laboratory actually make them and cool them down to see if they superconduct. All four did, at critical temperatures of about 6.5 K, 5.9 K, 3.5 K, and 2.5 K. Critical temperature is simply the temperature below which a material starts superconducting; the lower it is, the harder and more expensive it is to use in practice.

That last detail is the honest caveat here. Every one of these four materials only superconducts at extremely cold temperatures - all under about 7 K, which is colder than -266 C (-447 F) and close to absolute zero. None of them are the long-sought room-temperature superconductors that would let power grids, MRI machines, or maglev trains run without the bulky refrigeration equipment they currently need. The real news is not the materials themselves, which are unlikely to be used anywhere, but the method: an AI agent proposed genuinely new candidates, and independent physical experiments backed it up. That closes a loop - AI prediction to lab confirmation - that AI-for-science projects have promised for years but rarely delivered end to end.

As the team put it in describing their approach, the project is "Agentic Fusion of Large Atomic and Language Models to Accelerate Superconductor Discovery." That framing - fusion of a reasoning model with a physics model, working as an agent rather than a one-shot predictor - is likely to show up again well beyond superconductors, in any materials-discovery problem where screening millions of candidates by hand simply is not feasible. For more on how these agent loops work generally, see our explainer on /learn/ai-agents.html, and for another case of AI directly driving physical lab work, see /news/robots-run-experiments-themselves.html.

Primary source, verified: read the paper → (arXiv 2604.23758)

Ground Truth. (2026, July 4). An AI Agent Screened 2.4 Million Crystals and Found Four New Superconductors. Ground Truth. https://groundtruth.day/news/an-ai-agent-found-four-new-superconductors.html

@misc{groundtruth:an-ai-agent-found-four-new-superconductors,
  title  = {An AI Agent Screened 2.4 Million Crystals and Found Four New Superconductors},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/an-ai-agent-found-four-new-superconductors.html}
}

Topics: Alibaba · AI for science · materials · superconductors · agents