AI Framework Discovers New Carbon Structures Harder Than Diamond

Researchers at Xi'an Jiaotong University have developed a sophisticated AI-driven framework capable of methodically identifying new, stable forms of solid carbon. By integrating a Large Language Model (LLM) with a physics-based validation system, the team has successfully uncovered several carbon allotropes with exotic properties, including a superhard material that calculations suggest may exceed the hardness of diamond.

Expanding the Carbon Landscape

Solid carbon is defined by its unique bonding behavior, where atoms can link in linear (sp), trigonal planar (sp2), or tetrahedral (sp3) configurations. While these bonds can theoretically combine in near-limitless ways to form diverse structures, the vast majority of these potential allotropes remain undiscovered. Conventional search methods often struggle to navigate this complex configurational space, as the computational requirements for verifying stability become prohibitive.
To overcome this, Zhibin Gao and his colleagues utilized CrystaLLM, an LLM designed to model crystal structures. While CrystaLLM can generate a wide array of plausible carbon forms, it previously lacked the ability to verify their stability. The new framework addresses this by incorporating a "hybridization Shannon entropy" descriptor, which quantifies structural complexity and guides the AI to explore the underexplored space where sp, sp2, and sp3 bonds coexist.

Breakthrough Material Properties

The framework’s iterative, closed-loop process allowed the researchers to screen thousands of candidates, resulting in the discovery of materials with unprecedented characteristics. Among these is a dense, sp3-dominant network that exhibits extreme hardness. Other findings include a material with direction-dependent thermal conductivity and low shear stiffness, as well as a C12 phase that uniquely combines metallic conductivity with a negative Poisson's ratio, causing the material to expand perpendicularly when stretched.
Beyond these structural breakthroughs, the team’s electronic calculations indicate that certain yne-diamond phases function as narrow-bandgap semiconductors. These properties suggest potential applications in infrared technology and thermoelectric devices.

Pathways to Synthesis

The research team evaluated the practical viability of these structures, finding that their stability is comparable to existing carbon materials like fullerenes. The researchers suggest that these new forms could be realized in the laboratory through established chemical synthesis methods or by subjecting starting materials to high-pressure compression.
This scalable framework provides a new, efficient tool for materials science. By linking hybridization Shannon entropy directly to physical behavior, the approach serves as a predictive guide for navigating the topological landscape of carbon. According to Gao, the framework is also extensible to other elemental or multi-component systems, offering a path toward the inverse design of next-generation functional materials.
The study, titled "LLM-driven discovery for carbon allotropes with bond-network entropy," was published in Applied Physics Letters.