ElementsClaw: An Intelligent Material Discovery Agent Fusing Atomic Large Models and Language Models

The research team developed the ElementsClaw intelligent agent framework, which fuses large atomic models (LAM) with large language models to address the limitation of existing AI models operating in isolation, enabling autonomous execution of the entire material discovery process. The framework has achieved significant results in the superconductor field: it successfully guided the synthesis of 4 new superconductors and identified 68,000 high-confidence superconducting candidates through large-scale high-throughput screening.

The discovery of new materials is crucial for energy transition and quantum technology, but the traditional process is slow, costly, and relies on trial and error. Deep learning has revolutionized material discovery: predictive models can learn patterns, and generative models can propose candidates. However, existing AI models operate in isolation and require alternating human-machine decision-making, limiting their potential.

The core of ElementsClaw is the fusion of LAM and LLM: LAM handles atomic-scale numerical calculations (such as density functional theory calculations and crystal structure optimization), while LLM takes charge of semantic reasoning and decision coordination; the intelligent agent layer dynamically arranges tools to adapt to different scenarios. Technical innovations include toolized LAM design, closed-loop human-machine interaction, multi-objective optimization capabilities, and knowledge integration and transfer.

ElementsClaw has been successfully validated in the superconductor field: it guided the synthesis of 4 new superconductors (e.g., Zr3ScRe8 with a transition temperature of 6.8K); it evaluated 2.4 million stable crystal structures in 28 GPU hours and screened out 68,000 high-confidence candidates; the results are based on rigorous physical calculations, avoiding the "hallucination" problem of pure data-driven approaches.

ElementsClaw marks the shift of AI-driven materials science from isolated processes to an integrated model: accelerating the discovery cycle (from years to months/weeks), lowering research barriers (encapsulating computational complexity), expanding the exploration space (covering a broader chemical space), and promoting interdisciplinary integration (bridging atomic and language models).

ElementsClaw has limitations: high computational cost, bottlenecks in experimental validation, and generalization capabilities that need verification across more material categories. Future outlook: enhancing the capabilities of LAM and LLM, optimizing the agent architecture, and it is expected to play a role in more scientific fields.