ElementsClaw: Agentic Fusion of Large Atomic and Language Models to Accelerate Superconductor Discovery

Agentic AI is emerging as a powerful paradigm for accelerating materials discovery beyond standalone prediction and generation. Here we will present our recent work on ElementsClaw, an agentic framework that integrates large atomic models with large language models to address the decision problem in AI-driven materials discovery. At the atomic-scale modeling side, we develop Elements, a 1-billion-parameter foundation model pretrained on 125 million molecular and crystal structures, and finetune it into a suite of specialized tools for superconducting critical temperature prediction, superconductivity classification, thermodynamic stability evaluation, and crystal generation. At the reasoning side, ElementsClaw leverages large language models to integrate literature evidence, orchestrate tools, construct self-evolving workflows, and identify experimentally viable target systems.x000D Applied to superconductor discovery, ElementsClaw rediscovers 66 experimentally verified superconductors missing from the standard SuperCon3D database, and scales to 2.4 million equilibrium crystals to identify 68,000 high-confidence candidates in only 28 GPU hours. Guided by the agent’s reasoning, we experimentally synthesize and verify four new superconductors: motif-guided Zr3ScRe8, de novo generated HfZrRe4, structurally reinterpreted Zr4VRe7, and database-latent Hf21Re25. More broadly, this work demonstrates how knowledge-integrated, autonomously orchestrated, and experimentally grounded AI systems can bridge large-scale candidate generation with scientific decision-making, offering a general strategy for closed-loop materials discovery beyond superconductors.