Methods for predicting or discovering materials with specific thermal conductivity properties.Thermal conductivity prediction refers to computational and machine-learning methods for estimating or discovering materials with desired heat-transport properties, such as ultra-low or ultra-high thermal conductivity. This is crucial for applications like thermoelectric energy conversion, thermal barrier coatings, and heat management in electronics. The challenge lies in the fact that thermal conductivity depends on complex atomic-scale interactions — phonon scattering, crystal structure, defects, and composition — making it difficult to predict from simple rules.
The field emerged from early empirical correlations in the 20th century, but the modern era began in the 2000s with high-throughput computational screening. The use of deep learning for thermal conductivity, as exemplified by models like KappaFormer, is a development of the 2020s, driven by the availability of large materials databases and advances in graph neural networks.
Modern approaches combine first-principles calculations (like density functional theory) with machine learning models trained on large databases of known materials. For example, neural networks can learn to map atomic structures and compositions to thermal conductivity values, then screen millions of candidate compounds in silico. Key physical insights, such as the role of anharmonicity, phonon group velocity, and scattering rates, are encoded in the features used by these models. The best predictions often come from hybrid methods that use physics-based descriptors alongside data-driven regression.
These methods are used to discover new thermoelectric materials for waste-heat recovery, identify efficient thermal insulators for aerospace, and optimize heat spreaders for microelectronics.
Machine learning models have predicted materials with thermal conductivity lower than that of air, which could revolutionize insulation if synthesized.
| k | variable | thermal conductivity — thermal conductivity watt per metre per kelvin |
| C | variable | heat capacity per unit volume — volumetric heat capacity joule per cubic metre per kelvin |
| v | variable | average particle speed — speed metre per second |
| l | variable | mean free path — mean free path metre |
| m_eff | variable | Effective mass (temperature-dependent) — mass kilogram |
| T | variable | Temperature — temperature kelvin |
| m_0 | variable | Bare mass at zero temperature — mass kilogram |
| c | variable | Coefficient depending on coupling constants — coefficient kilogram per square kelvin |
| U_d | variable | Depletion attraction energy — depletion energy joule |
| k_B | variable | Boltzmann constant — boltzmann constant joule per kelvin |
| T | variable | Absolute temperature — temperature kelvin |
| Π_p | variable | Osmotic pressure of depletant particles — osmotic pressure pascal |
| V_o | variable | Overlap volume of depletion zones — overlap volume cubic metre |
