Kuaikeji, July 3rd news: Superconducting materials boast immense potential in applications spanning power transmission, maglev trains, and quantum devices. Yet, for over a century, the lack of clarity in underlying mechanisms has made new material development reliant on inefficient trial-and-error methods. The globally recognized authoritative superconducting database, SuperCon, has amassed just around 2,000 materials over decades.

To tackle this challenge, Alibaba DAMO Academy, in partnership with Renmin University of China and the University of the Chinese Academy of Sciences, has unveiled ElementsClaw—an AI agent meticulously crafted for new material R&D. By harnessing AI-driven reasoning and the capacity to autonomously devise experiments, it revolutionizes the conventional trial-and-error R&D paradigm.

This agent employs a hybrid architecture combining a "specialized atomic large model with a general intelligence framework." It has trained a 1B-parameter atomic foundation model using 125 million molecular and crystal datasets, achieving remarkable accuracy in discerning superconducting properties. The error margin in predicting superconducting critical temperatures is controlled within 1K.

Simultaneously, it can autonomously search academic literature, organize research leads, and self-iterate, fully automating the entire material screening and experimental design process.

Its computational power consumption is remarkably low, requiring only 28 GPU hours to screen 2.4 million crystal structures, thereby identifying 68,000 candidate substances with superconducting potential.

The research team selected four categories from the vast array of candidates for physical synthesis and experimental validation. The four novel materials originate from diverse sources: HfRe, previously overlooked in the database; ZrVRe, confirmed after rectifying its original structural errors; HfZrRe, a completely new crystal designed by AI from the ground up; and ZrScRe, derived through analogous structural deduction. Among them, the highest critical temperature reaches 6.5K.

These are also the first batch of superconducting materials independently discovered by AI in China and experimentally verified, vividly showcasing the agent's practical value in the realm of cutting-edge materials.

The research team has made the entire database of 2.4 million stable crystal predictions publicly accessible, allowing researchers worldwide to freely access and utilize it.

Experts note that this technology is not confined to superconducting R&D but can also be extended to the exploration of various new materials, such as solid-state battery electrolytes, catalysis, and thermoelectrics, in the future. The relevant research paper has been published on arXiv, and the research team will also present this AI material discovery system at the ICML conference.