A new machine learning framework could revolutionize the efficiency of energy generation systems by predicting the transfer of heat through semiconductors and insulators with unprecedented speed and accuracy.
On average, about 70 percent of the energy generated in the global energy system is lost as waste heat, and addressing this inefficiency has been a major historical challenge for electrical engineers so that each unit of electricity generated can better supply consumers' energy needs, saving energy and potentially significantly reducing carbon emissions.
However, solving this problem requires understanding the thermal properties of materials, a complex task due to the behavior of phonons, the subatomic particles that carry heat. The phonon dispersion relation (PDR), which describes the relationship between the energy and momentum of phonons within a material's crystal structure, is particularly difficult to model.
Now, a team led by engineers at the Massachusetts Institute of Technology (MIT) is tackling this challenge with a new machine learning framework that predicts PDR up to 1,000 times faster than existing AI techniques and up to 1 million times faster than traditional methods.
The technique is described in a new paper published in the journal Nature. Nature Computational Scienceenables greater efficiency in power generation systems and microelectronics designs, where thermal management has traditionally been a major bottleneck.
“Phonons are responsible for thermal loss, but their characterization has been notoriously difficult both computationally and experimentally,” Minda Li, an MIT associate professor of nuclear science and engineering and lead author of the paper, said in an MIT news release.
Heat-carrying phonons are difficult to predict due to their wide frequency range and variable travel speed. Traditional machine learning models such as graph neural networks (GNNs) struggle with the high-dimensional nature of phonon dispersion relations. To overcome this, researchers developed virtual node graph neural networks (VGNNs), which introduce flexible virtual nodes in a fixed crystal structure, allowing the model to adapt and efficiently predict phonon behavior.
“This method is very efficient in coding; we just generate a few more nodes in the GNN. Their physical location doesn't matter, and the real nodes don't even know that the virtual nodes exist,” said Abijatmedi Chotrattanapituk, a graduate student at MIT and co-author of the paper.
The study claims that VGNN can estimate phonon dispersion relations quickly and with slightly higher accuracy in predicting the heat capacity of materials. This efficiency could allow the phonon dispersion relations of thousands of materials to be calculated on a personal computer within seconds, accelerating the discovery of materials with superior thermal properties.
Potential applications for new heat-resistant materials
Going forward, the researchers aim to improve the technique by increasing the sensitivity of the virtual nodes to capture subtle changes that affect the phonon structure.
“The nodes in the graph can be anything,” Li says. “Virtual nodes are a very general approach that can be used to predict many high-dimensional quantities.”
This innovative framework not only promises to improve energy efficiency but also opens new avenues for research into optical and magnetic properties, potentially transforming multiple areas of materials science.
About the Editor
John Leffler John is a writer and programmer living in New York City. He writes about computers, gadgets, games, VR/AR, and related consumer tech. You can find him on Twitter at @thisdotjohn.
