- ElementsClaw identified 68,000 potential superconducting materials, with four newly synthesized and experimentally verified
- Researchers say the system automates literature review, experiment design and materials screening, replacing trial-and-error with AI-driven discovery
Alibaba’s research unit DAMO Academy and researchers from Renmin University of China and University of Chinese Academy of Sciences on July 3 unveiled an AI agent designed to accelerate the discovery of superconducting materials.
This agent reportedly identified 68,000 promising candidates, leading to the experimental verification of four previously unknown superconductors.
The system, called ElementsClaw, is the first AI agent developed specifically for superconducting materials discovery, according to the research team.
Automating workflows
Researchers said it combines large-scale materials prediction with autonomous scientific reasoning, allowing it to search literature, evaluate synthesis feasibility and design experimental workflows in addition to predicting superconducting properties.
Superconductors are materials that lose all electrical resistance below a critical temperature while expelling magnetic fields, making them attractive for applications including power transmission, magnetic levitation and advanced computing.
Yet discovering new superconductors has traditionally relied on decades of labor-intensive experimentation because the underlying physics remains poorly understood.
The widely used international SuperCon database, for example, contains only about 2,000 known superconducting materials after decades of research.
Hybrid architecture
To address this bottleneck, the team developed ElementsClaw using a hybrid architecture that combines a domain-specific foundation model with a general AI agent framework.
At its core is Elements, a one-billion-parameter atomic foundation model pretrained on a database containing 125 million molecular and crystal structures.
According to the researchers, the model achieved an AUC score of 0.996 in identifying superconducting candidates.
AUC, which stands for Area Under the ROC Curve, is a measure of a model’s ability to distinguish between two classes. The closer to 1.0, the better it is at separating the “good” cases from the “bad” ones.
Additionally, ElementsClaw predicted critical temperatures with an average error of less than one kelvin.
Beyond forecast, the AI agent automates the full materials discovery workflow, including tool generation, literature review, experimental planning and iterative refinement as new scientific evidence becomes available.
Validating four candidates
Using just 28 GPU-hours, ElementsClaw screened 2.4 million crystal structures and identified approximately 68,000 candidate superconductors.
Researchers have since synthesized and experimentally verified four materials, including one generated entirely by the AI and others identified by correcting existing database records or extrapolating from related crystal structures.
The highest-performing material demonstrated superconductivity at a critical temperature of 6.5 kelvin.
Rong Yu, head of AI for Science at Alibaba DAMO Academy, said the four experimentally validated materials represent the first batch of superconductors discovered with the assistance of an autonomous AI agent, demonstrating the framework’s potential to accelerate materials research.
“DAMO Academy has open-sourced the full database of 2.4 million stable crystal structures predicted by ElementsClaw for free use by researchers,” Rong added.
Huang Wenbing, associate professor at Renmin University’s Gaoling School of Artificial Intelligence, said the AI agent could also be applied to the discovery of new materials such as solid-state battery electrolytes, heterogeneous catalysts and thermoelectric materials.
