Accelerating Discovery of Extreme Lattice Thermal Conductivity by Crystal Attention Graph Neural Network (CATGNN) Using Chemical Bonding Intuitive Descriptors

Heat moves through solid materials in a way that is often invisible but fundamentally important to how our technology works. In crystals, which are solids with atoms arranged in a perfect, repeating pattern, heat travels not as a fluid but as vibrations of the atoms themselves. These vibrations, known as phonons, carry energy from one end of a material to the other. For some applications, such as cooling down a powerful computer chip or a high-speed laser, engineers need materials that let this heat race through as fast as possible. For other tasks, like keeping a thermoelectric generator efficient or insulating a building, they need materials that trap the heat, slowing the vibrations to a crawl. The challenge has always been finding new materials with these extreme properties. While scientists have long known that heavy atoms or weak bonds tend to slow heat down, and light, strong bonds tend to speed it up, these rules are often too vague to predict exactly how a specific new crystal will behave.

A team of researchers has now developed a much sharper way to see inside these materials, moving beyond simple rules of thumb to a detailed understanding of how atoms hold hands. By analyzing the electronic glue that binds atoms together, they created two new ways to measure the strength of these connections. They found that these measurements act as a reliable compass for predicting whether a crystal will be a thermal superhighway or a thermal dead end. Using this insight, they trained a sophisticated computer model to scan nearly 200,000 known crystal structures. The search was successful: the model identified dozens of new materials that are dynamically stable and possess extreme thermal properties, including some that conduct heat less than 2 watts per meter-kelvin and others that conduct it over 100 watts per meter-kelvin. This work offers a rapid, reliable path to discovering the next generation of materials for thermal management, from better insulators to more efficient energy converters.

The journey began with a realization that the old ways of guessing thermal behavior were hitting a wall. For decades, scientists have relied on factors like the average weight of the atoms in a material or the general simplicity of its structure to estimate how well it conducts heat. While these factors matter, they often fail to distinguish between materials that look similar on paper but behave very differently in reality. To solve this, the researchers turned their attention to the chemical bonds themselves. They focused on two specific electronic signatures: one that measures how much energy is released when atoms form a bond, and another that quantifies how covalent, or shared, that bond is. They realized that if they normalized these values—adjusting them to account for the number of atoms and bonds in the crystal—they could create a universal scale for bond strength that works across thousands of different materials.

These new descriptors, which the team calls normalized negative integrated crystal orbital Hamilton population and normalized integrated crystal orbital bond index, turned out to be incredibly powerful. When the researchers plotted these values against the actual thermal conductivity of over 4,500 materials, a clear pattern emerged. Materials with high values on these scales, indicating strong, deep electronic bonds, consistently allowed heat to flow freely. Conversely, materials with low values, indicating weak or shallow bonds, trapped the heat. Crucially, this relationship held true even when the materials had the same average atomic mass, a factor that previous models relied on heavily. The new descriptors also revealed a direct link to how much the atoms wobble when heated. Strong bonds kept the atoms steady, while weak bonds allowed them to rattle wildly, a motion that scatters heat and slows it down. This connection provided a physical explanation for why some materials are thermal insulators and others are conductors, rooted directly in the electronic structure of the atoms.

To put this discovery to work, the team built a new type of artificial intelligence model called a Crystal Attention Graph Neural Network. Think of this model as a highly trained eye that can look at the arrangement of atoms in a crystal and instantly "feel" the strength of the bonds holding them together, without needing to run the slow, expensive computer simulations that usually take days or weeks. The researchers first taught this model using the data from the 4,500 materials they had already analyzed. Once the model learned the relationship between the crystal structure and the new bond-strength descriptors, they unleashed it on a massive database containing nearly 200,000 untested crystal structures. The model quickly predicted the bond strength for each one, flagging the most promising candidates for extreme thermal behavior.

From this massive digital sweep, the researchers selected about 900 materials that showed the most extreme bond strengths—either very weak or very strong. They then subjected these candidates to a rigorous two-step verification process. First, they used a separate AI model to check if the crystals were stable, meaning they wouldn't fall apart or vibrate themselves into chaos. About 80 percent of the candidates passed this initial stability test. For those that remained, the team performed detailed first-principles calculations, the gold standard of computational physics, to confirm their thermal conductivity. The results were striking. The model had successfully identified 106 stable materials with ultralow thermal conductivity, with 68 percent of them conducting less than 2 watts per meter-kelvin. On the other end of the spectrum, the search yielded 13 stable materials with ultrahigh thermal conductivity, all exceeding 100 watts per meter-kelvin.

The study also challenged some long-held beliefs about what causes materials to block heat. Previous research had suggested that the presence of specific "antibonding" electronic states was the primary culprit behind low thermal conductivity. However, the new analysis showed that this was not a universal rule. The researchers found materials with these antibonding states that still conducted heat well, and materials without them that were excellent insulators. Instead, the new bond-strength descriptors proved to be the consistent predictor. This distinction is vital because it means scientists can no longer rely on a single electronic feature to screen for thermal properties; they must look at the overall strength of the chemical bonding.

The implications of this work extend beyond just finding new materials. By establishing a clear link between electronic bonding and thermal transport, the researchers have provided a set of intuitive principles that can guide the design of future materials. The team demonstrated that their approach works for a wide variety of chemical families, including those with heavy elements like lead and bismuth, as well as lighter elements like carbon and boron. The successful identification of specific compounds, such as various forms of carbon, boron nitride, and complex alloys containing alkali metals, proves that the method is robust. These findings suggest that the path to discovering the next generation of thermal materials is no longer a game of chance or a slow process of trial and error, but a targeted search guided by the fundamental physics of how atoms connect.