Alibaba's AI agent (Elements Claw) Discovers 4 New Superconductors

Alibaba Group Holding's research arm, DAMO Academy, announced Friday the launch of an artificial intelligence agent that independently identified four previously unknown superconducting materials. The system, ElementsClaw, screened 2.4 million crystal structures in 28 hours of graphics processing unit computing time. The development team successfully synthesized and experimentally verified all four predicted compounds in laboratory tests.

Superconducting materials conduct electricity with zero resistance and expel external magnetic fields when cooled to specific low temperatures. These properties make them valuable for power transmission grids, quantum computing hardware, and high-speed magnetic levitation trains. Traditional methods for discovering these materials rely on slow, trial-and-error experiments because scientists still lack a complete theoretical framework to predict superconductivity.

The project represents a joint effort between DAMO Academy, Renmin University of China, and the University of Chinese Academy of Sciences.

The international superconductor database, SuperCon, currently contains only about 2,000 materials accumulated over several decades. The research team stated that ElementsClaw shortens this discovery timeline by mimicking the full workflow of human materials scientists.

The software architecture combines a specialized foundation model with general-purpose automation capabilities. The team pre-trained a one-billion-parameter atomic foundation model called Elements using a database of 125 million molecular and crystal structures. This primary model achieved a 0.996 Area Under the Curve score in determining whether a material exhibits superconductivity, limiting its average critical temperature prediction error to within one Kelvin.

ElementsClaw operates an automated pipeline that reads scientific literature, assesses synthesis feasibility, estimates toxicity levels, and generates experimental protocols. The agent also extracts new insights from scientific papers to update its own evaluation rules automatically. Out of the 2.4 million crystal structures evaluated, the system flagged 68,000 candidate materials with superconducting potential.

The researchers selected the most viable candidates from this pool for physical testing. Laboratory experiments confirmed the superconductivity of four distinct chemical structures, recording critical temperatures up to 6.5 Kelvin. The final selection included HfZrRe4, which the AI designed from scratch, alongside Hf21Re25, Zr4VRe7, and Zr3ScRe8, which the system located by correcting database errors and analyzing similar structures.

Rong Yu, the head of the Scientific Intelligence department at DAMO Academy, detailed the trial validation in an official statement.

"These are the first batch of superconducting materials discovered and verified by an AI agent, preliminarily validating the potential of the AI agent framework in the field of material discovery."

DAMO Academy uploaded the entire database of 2.4 million stable crystal structures, including the specific superconductivity indicators and temperature predictions, to science.damo-academy.com. The institution made the entire dataset freely accessible to accelerate international research. Tech firms worldwide increasingly deploy automated models to scientific disciplines, moving past text chatbots to target physical engineering challenges.

Similar efforts by global technology firms include Google DeepMind's AlphaFold model for protein folding predictions and Microsoft's specialized systems for material design.

Academic contributors noted that materials science presents unique challenges compared to biological modeling due to the hundreds of chemical elements and complex bonding patterns involved. The ElementsClaw framework will expand next into searching for solid-state battery electrolytes and chemical catalysts.