AI Method Speeds Predictions of Materials’ Thermal Properties by 1000x

It is estimated that about 70% of the energy generated worldwide ends up as waste heat. Better predictions of how heat moves through semiconductors and insulators could lead to more efficient power generation systems and reduce waste heat. However, predicting the thermal properties of materials is challenging due to the complexity of phonons, subatomic particles that carry heat.

New Machine-Learning Framework

A team of researchers from MIT and other institutions has developed a new machine-learning framework that predicts phonon dispersion relations up to 1,000 times faster than other AI-based techniques, and potentially 1 million times faster than traditional methods. This method offers comparable or even better accuracy, which could help engineers design energy generation systems that produce more power efficiently and develop faster microelectronics.

The Challenge of Phonons

Phonons are difficult to predict due to their wide frequency range and interactions that vary in speed. A material’s phonon dispersion relation, which shows the relationship between energy and momentum of phonons in its crystal structure, is crucial for understanding thermal properties. Traditional machine learning models for these predictions often become bogged down by high-precision calculations.

“Phonons are the culprit for the thermal loss, yet obtaining their properties is notoriously challenging, either computationally or experimentally,” says Mingda Li, associate professor of nuclear science and engineering at MIT.

Virtual Node Graph Neural Network (VGNN)

The new approach involves a virtual node graph neural network (VGNN), which adds flexible virtual nodes to the fixed crystal structure of the material. These virtual nodes represent phonons and enable the neural network output to vary in size, bypassing many complex calculations and making the method more efficient. This innovation allows for rapid estimation of phonon dispersion relations, significantly speeding up the prediction process.

Practical Applications and Future Directions

The VGNN can predict phonon dispersion relations for thousands of materials in seconds using a personal computer, enabling researchers to explore a broader range of materials with desirable thermal properties. The technique could also be adapted for other challenging properties, such as optical and magnetic characteristics.

Olivier Delaire, associate professor at Duke University, who was not involved in the work says, “I find that the level of acceleration in predicting complex phonon properties is amazing, several orders of magnitude faster than a state-of-the-art universal machine-learning interatomic potential. Impressively, the advanced neural net captures fine features and obeys physical rules. There is great potential to expand the model to describe other important material properties: Electronic, optical, and magnetic spectra and band structures come to mind.”

Collaboration and Support

The research team includes members from MIT, Argonne National Laboratory, Harvard University, the University of South Carolina, Emory University, the University of California at Santa Barbara, and Oak Ridge National Laboratory. The work is supported by the U.S. Department of Energy, the National Science Foundation, a Mathworks Fellowship, a Sow-Hsin Chen Fellowship, the Harvard Quantum Initiative, and Oak Ridge National Laboratory.