The Century-Old "Trial-and-Cooking" Problem in Superconductor Discovery
Since Heike Kamerlingh Onnes discovered superconductivity in mercury in 1911, humanity has found only about 2,000 superconducting materials, according to the international SuperCon database. Researcher Jin Shifeng of the University of Chinese Academy of Sciences describes the process as "stir-fry-style research": no recipe exists, so scientists repeatedly adjust elemental ratios, with success rates so low that dozens of experiments may yield only one success.
On July 3, 2026, Alibaba DAMO Academy—jointly with Renmin University of China and the University of Chinese Academy of Sciences—released Elements Claw, billed as the world's first AI agent dedicated to superconducting material discovery. In just 28 GPU hours (roughly one graphics card running for a day), it screened all 2.4 million known stable crystal structures and identified 68,000 candidates with superconducting potential—expanding the candidate pool more than 30-fold beyond a century of human accumulation.
> Note on Tc: The critical temperature (Tc) is the upper temperature limit at which a material enters the superconducting state. Higher Tc means lower engineering cost. High-Tc materials such as cuprates, nickelates, and magnesium diboride were almost all discovered by experimental accident. Elements Claw's core capability is predicting Tc with ±1 K error.
The "Specialist-Generalist Fusion" Architecture
Elements Claw combines two models with complementary roles:
- Specialist brain: Elements, a 1-billion-parameter atomic foundation model handling structure prediction, energy estimation, and Tc prediction.
- Generalist brain: An LLM responsible for literature reading, synthesis feasibility assessment, interdisciplinary knowledge integration, and experiment design.
- Which of the remaining ~67,900 candidates enter experimental synthesis (a key test of conversion rate)
- Whether the framework transfers to solid-state batteries, catalysts, and thermoelectrics
- Whether the DAMO–university collaboration model scales into national-level AI materials infrastructure
- Whether Tc of the four new materials can be raised via doping, pressure, or nanostructuring
- Whether Google DeepMind, Microsoft Research, or Meta AI replicate the methodology
- Liability questions when AI-designed materials exhibit safety issues
Key specifications:
| Dimension | Value | |---|---| | Pretraining corpus | 125 million molecular/crystal structures | | Parameters | 1 billion (1B) | | Initial crystal library | 2.4 million stable structures | | Candidate materials | 68,000 | | Screening compute | 28 GPU hours | | Superconductivity classification AUC | 0.996 | | Tc prediction error | ±1 K |
The division of labor is deliberate: the specialist model handles numerical precision while the generalist model provides knowledge breadth. A single model attempting both typically underperforms a two-model split.
Four Experimentally Verified New Materials
AI candidates are only the starting line; the real test is synthesis and verification. Elements Claw delivered four new superconducting materials:
| Material | Origin | Key feature | |---|---|---| | HfZrRe4 | Designed from scratch by AI | Zr-Hf-Re alloy, Tc = 6.5 K | | Hf21Re25 | Correcting/reanalyzing existing data | Hafnium-rhenium compound | | Zr4VRe7 | Correcting/reanalyzing existing data | Zirconium-vanadium-rhenium compound | | Zr3ScRe8 | Correcting/reanalyzing existing data | Zirconium-scandium-rhenium compound |
HfZrRe4 is the key breakthrough: never synthesized before, it was an entirely new combination proposed autonomously by the AI from the elemental library. At 6.5 K, it remains superconducting above the liquid-helium temperature (4.2 K), lowering the barrier for engineering use—though far from room-temperature superconductivity. The four materials went from AI candidate to experimental verification in under one month, versus a traditional 2–5 year timeline.
A Reusable Materials-Discovery Pipeline
DAMO Academy's scientific intelligence lead Rong Yu summarized the significance as validating the potential of AI agents in materials discovery. The division of roles:
| Institution | Role | |---|---| | Alibaba DAMO Academy | Elements model architecture, fusion framework, compute | | Renmin University of China | LLM generalist brain, synthesis feasibility, literature integration | | UCAS (Jin Shifeng's team) | Experimental synthesis, Tc measurement, physical verification |
DAMO Academy has open-sourced the 2.4-million-crystal dataset. Professor Huang Wenbing of Renmin University believes the framework can be reused for solid-state battery materials, catalysts, thermoelectric materials, and other inorganic functional materials—making Elements Claw a general materials-discovery framework, not just a superconductor finder.
Two Open Questions from the Scientific Community
1. Explainability: The AI produced 68,000 candidates and 4 verified materials, but did not explain *why* HfZrRe4 superconducts. It works via data pattern matching rather than a physical causal chain—useful, but not yet explanatory.
2. Practical distance to room-temperature Tc: All four materials have Tc at or below 6.5 K. The leading nickelate and hydride systems remain below 200 K, and even high-pressure hydrides (e.g., LaH10, ~250 K) require over 200 GPa. Room-temperature superconductivity (Tc ≥ 300 K) remains the "holy grail."
Toward "AI-Autonomous Scientific Research"
Elements Claw exemplifies an emerging paradigm with three traits:
1. Autonomous task definition — the LLM reads literature, judges which directions are worth exploring, and designs experiments itself. 2. Autonomous tool use — calling DFT software, databases, and lab equipment APIs to automate the pipeline. 3. Autonomous iteration — learning from experimental feedback, turning failures into training data.
Currently Elements Claw operates in a "dual-leadership" mode: AI leads screening and design, humans lead experimental verification.
What to Watch Over the Next 6–12 Months
Conclusion
Elements Claw transformed superconductor exploration from luck-based trial and error into an AI-driven pipeline: 68,000 candidates identified in 28 GPU hours, a 30-fold expansion of the search space, and four verified new materials in under a month. Its deeper significance is architectural—a reusable specialist-generalist framework with closed-loop experimental feedback, applicable wherever "trial-and-error plus screening plus verification" is required.
References
1. Guangming Net via Toutiao, *AI as superconductor treasure hunter unlocks 4 new materials*, 2026-08-27 — https://www.toutiao.com/article/7678642444405080582/ 2. AI DAMN, *AI Agent 'Elements Claw' Cracks Superconducting Material Discovery*, 2026 — https://ai-damn.com/ai-agent-elements-claw-cracks-superconducting-material-discovery-1783292505865 3. Auto Info Net, *Discovering new materials: AI enters autopilot mode*, 2026 — https://e.mffb.com.cn/a/534080.html 4. Alibaba DAMO Academy official Elements Claw materials (open-sourced 2.4M crystal dataset) 5. Jin Shifeng (UCAS), "stir-fry-style research" analogy 6. Elements model official data: 125M structures / 1B parameters / AUC 0.996 / Tc ±1 K 7. SuperCon database baseline of ~2,000 superconducting materials