An international team of researchers, including members from MIT (Massachusetts Institute of Technology), has developed a machine learning-based approach to predict the thermal properties of materials. This understanding could help improve energy efficiency in power generation systems and microelectronics.
The research focuses on phonons – subatomic particles that carry heat. Properties of these particles affect a material’s heat characteristics, a relationship referred to as the ‘phonon dispersion relation’. To date, this relation has been difficult to predict or obtain using conventional methods either computationally or experimentally, proving a significant challenge for researchers in the field.
The team, led by MIT’s Associate Professor of Nuclear Science and Engineering, Mingda Li, adapted machine learning models known as graph neural networks to create a model which they named the ‘virtual node graph neural network’ (VGNN). By using virtual nodes to represent phonons, the VGNN vastly improves the efficiency of the computational process, without sacrificing accuracy. In tests, the VGNN predicted phonon dispersion relations up to 1,000 times faster than other AI-based techniques, and 1 million times faster than traditional non-AI methods.
Adopting this approach could significantly aid scientists and engineers in the design and development of more efficient energy systems and electronic devices. In a broader context, the researchers believe that the VGNN model could also be adapted to predict other challenging material properties such as optical or magnetic characteristics.
This innovative framework could also be used in the future to expand the search for materials with desirable thermal properties, due to its ability to quickly process a large number of materials. For instance, the researchers were able to calculate the phonon dispersion relations for several thousand materials in just a few seconds using a personal computer.
The research, which has been published in the journal Nature Computational Science, received funding from the U.S. Department of Energy, National Science Foundation, and the Mathworks Fellowship, along with other sources. Moving forward, the team plans to refine the VGNN further to capture subtle changes that can affect the phonon structure of materials.
