Alibaba DAMO Academy Unveils AI Agent ElementsClaw, Discovers 4 New Superconductors in 28 GPU Hours

The century-long race for superconductor discovery has reached a new milestone with AI entering the arena. Alibaba DAMO Academy, in collaboration with Renmin University's Gaoling School of Artificial Intelligence and the University of Chinese Academy of Sciences, unveiled ElementsClaw, an AI agent purpose-built for discovering superconducting materials.

ElementsClaw screened 2.4 million known stable crystal structures in just 28 GPU hours, predicting 68,000 of them could be superconductors. From these candidates, the team experimentally synthesized and validated 4 entirely new superconductors unknown to humanity. By comparison, the mainstream SuperCon database contains only around 2,000 superconducting materials discovered over more than a century of human research.

The agent's architecture follows a general-specialist fusion design. At its core is Elements, a 1-billion-parameter geometric deep graph neural network pretrained on 125 million molecular and crystal structures. On top of this, multiple specialized models handle different tasks: Elements-T predicts superconducting critical temperature with a mean absolute error of just 0.99K, Elements-C classifies whether a material is superconducting, Elements-E assesses stability, and Elements-G generates novel crystal structures. A large language model ties these together by reading academic papers, querying databases, and evaluating synthesizability.

The four newly discovered superconductors each came through different discovery pathways. One, Hf21Re25, was already in theoretical databases but had never been synthesized. Another, Zr4VRe7, had been mischaracterized due to an incorrect crystal structure assignment. A third, HfZrRe4, existed in no known database and was generated from scratch. The fourth, Zr3ScRe8, was found by the AI detecting structural patterns from its earlier discoveries.

Unlike single-point prediction models such as DeepMind's GNoME or Microsoft's MatterGen, ElementsClaw functions as a complete autonomous agent. It can search literature, cross-reference databases, evaluate synthesizability, design experiments, and even fine-tune its own models when encountering new data — capabilities that go far beyond simple property prediction.

DAMO Academy has released the full prediction database for 2.4 million stable crystals to the global research community at no cost. While the critical temperatures of the discovered superconductors (up to 6.5K) remain far from room-temperature superconductivity, the work demonstrates that AI agents can now close the full loop from computational prediction to experimental validation in materials science.