Alibaba DAMO Academy's Elements Claw: AI Agent Discovers Superconducting Materials From Scratch, 4 Synthesized and Verified

Published: 2026-07-03 11:27 (Beijing Time) Source: IT之家 (Alibaba DAMO Academy official account) Original article: https://www.ithome.com/0/972/089.htm Paper: https://arxiv.org/abs/2604.23758 Database: https://science.damo-academy.com/#/material

What Happened

On July 3, Alibaba DAMO Academy, in collaboration with Renmin University of China and the University of Chinese Academy of Sciences, released Elements Claw, described as the first AI agent for superconducting material discovery. The system predicted 68,000 candidate materials from 2.4 million crystal structures, and 4 entirely new materials have been synthesized and confirmed to exhibit superconductivity.

Core architecture: "specialized-general fusion".

• Specialized model: Elements, a 1B-parameter atomic foundation model pretrained on 125 million molecular/crystal structures, achieving AUC 0.996 for superconductivity classification and a mean critical-temperature prediction error below 1 K
• General agent framework: automates the full materials-screening pipeline — tool creation, workflow orchestration, and literature review — and can "self-evolve" after mining new clues from the literature

Output data:

• 2.4 million crystal structures screened in 28 GPU-hours
• 4 synthesized materials verified as superconducting:
• Hf₂₁Re₂₅ — a "missed" candidate recovered from existing databases (previously mislabeled)
• Zr₄VRe₇ — "rehabilitated" after correcting a database configuration error
• HfZrRe₄ — designed de novo by the AI
• Zr₃ScRe₈ — derived by analogy from similar structures, with the highest critical temperature at 6.5 K
• A database of 2.4 million stable crystals is fully open

Application extensions: Associate Professor Huang Wenbing of Renmin University noted potential uses in solid-state battery electrolytes, heterogeneous catalysts, and thermoelectric materials.

Deep Analysis

Elements Claw is not "AI accelerating materials discovery" — it is "AI independently completing materials discovery." The difference between these two phrasings is one word, but the meaning differs by a light-year.

1. The boundary from "AI-assisted science" to "AI-led science" has been crossed. In previous AI4Science landmarks — DeepMind's AlphaFold predicting protein structures, for example — AI always "provided candidate answers for human scientists." Elements Claw searches the literature itself, evaluates synthesis feasibility itself, designs experimental plans itself, and iterates on its own. HfZrRe₄ was designed from scratch by AI and then synthesized and verified by humans — empirical evidence of AI as the subject of discovery.

2. 1B parameters achieving 0.996 AUC — this is not brute force, but a victory of "data + architecture." Pretraining on 125 million structures plus specialized-general fusion (a specialized foundation model + a general agent framework) costs far less than training a 70B general model and fine-tuning it. Lesson for teams wanting to build large models: a vertical-domain 1B model + agent framework may outperform a fine-tuned general 70B.

3. Open-sourcing the database matters more than the paper. The team has made all 2.4 million predicted stable crystals fully open to researchers for free — the same ecosystem strategy as openly publishing model pricing or open-sourcing code libraries: short term, giving up a data moat; long term, building an ecosystem moat.

4. First empirical proof of AI agents "self-evolving" in the physical world. Elements Claw self-evolves after mining new clues from literature — a proactive agent loop, but running in materials science, one of the hardest domains: not self-evolution in chat, but altering real-world matter.

Why It Matters

1. AI for Science moves from "prediction" to "discovery." After AlphaFold, the next paradigm shift is "AI proposes hypotheses + AI designs experiments + AI verifies" — Elements Claw is the first complete industrial-scale demonstration of this path. 2. "Specialized-general fusion" may be a differentiated advantage for Chinese AI teams. A vertical 1B model + agent framework is a different playbook — compute constraints forcing architectural innovation. 3. AI agents' true home turf is the physical world, not the chat window. The real 2026 battleground for AI agents is not customer service or email — it is research, manufacturing, healthcare, and energy. 4. An engineering roadmap for materials/chemistry/biology teams. Pretrained 1B foundation model + agent framework + literature mining + experimental closed loop — this template can be replicated in any vertical scientific domain with known data and a known search space.

Risks and Open Questions

• 6.5 K critical temperature is far from high-temperature superconductivity. Cuprates and iron-based superconductors reach 100 K+; the HfZrRe₄ family are conventional BCS superconductors. A one-order-of-magnitude improvement in Tc is needed for real industrial value.
• No mention of interpretability. Elements Claw gives black-box predictions without explaining *why* a material superconducts, which may affect academic acceptance.
• "Stable" in the 2.4-million-crystal database means DFT-computationally stable, not synthetically stable. Theoretical stability ≠ lab synthesizability. How many of the ~68,000 remaining candidates can actually be synthesized is unknown.
• Boundary with DeepMind's GNoME (2.2 million stable structures, 2023) is unclear. The paper's title includes the word "Accelerate," suggesting Elements Claw may be closer to "superconductivity screening + experimental validation on top of GNoME-like work" than fully independent discovery.

In one sentence: Elements Claw is not "AI built another database" — it is the first time AI acted as the subject of discovery across the full materials-science pipeline. The formula it points to — 1B specialized model + agent framework + open database — may be more likely to produce real scientific breakthroughs than a 70B general model.