An MIT team develops a new AI method that can increase the speed of predicting m

It is estimated that about 70% of the world's energy is ultimately lost in the form of waste heat.

If scientists could more accurately predict how heat moves in semiconductors and insulators, they could design more efficient power generation systems. However, the thermal properties of materials, especially those related to a subatomic particle called phonons, are extremely complex to predict. Phonons carry heat energy, and some thermal properties of materials depend on a measurement method called the phonon dispersion relationship, which is difficult to obtain, let alone utilize in system design.

To address this challenge, researchers at MIT and their collaborators have rethought the entire problem. Their result is a brand-new machine learning framework that can predict the phonon dispersion relationship at a speed 1000 times faster than other artificial intelligence technologies, while maintaining equal or even higher accuracy. Compared with traditional non-AI methods, this new method can achieve a speed increase of up to 1 million times.

One of the main authors of the relevant paper on this technology is Mingda Li, an associate professor of nuclear science and engineering, who explained: "Phonons are the main culprits of heat loss, but it is very difficult to obtain their properties either computationally or experimentally."

Advertisement

Also involved in this paper are co-first authors Ryotaro Okabe, a chemistry graduate student; Abhijatmedhi Chotrattanapituk, a graduate student in electrical engineering and computer science; Tommi Jaakkola, MIT Thomas Siebel Professor of Electrical Engineering and Computer Science; and other researchers from MIT, Argonne National Laboratory, Harvard University, University of South Carolina, Emory University, University of California, Santa Barbara, and Oak Ridge National Laboratory. This research was published in Nature Computational Science.Predicting Phonons

The frequency range of heat-carrying phonons is extremely broad, and the interactions between particles and the speeds of their propagation vary greatly, making predictions difficult.

The phonon dispersion relation refers to the relationship between the energy and momentum of phonons within their crystal structure. For many years, researchers have been attempting to use machine learning to predict the phonon dispersion relation, but due to the involvement of a large number of high-precision calculations, the models are very slow to process.

"If you have 100 CPUs and spend several weeks, you might be able to calculate the phonon dispersion relation of a material. The entire research community really needs a more efficient method to do this," said Okabe.Scientists commonly use a type of machine learning model known as Graph Neural Networks (GNN). GNNs convert the atomic structure of materials into a lattice graph composed of multiple nodes, where these nodes represent atoms, and the edges represent the bonds between atoms.

While GNNs perform well in calculating many physical quantities such as magnetization or polarization, they lack flexibility when it comes to high-dimensional quantities like phonon dispersion relations. Phonons can move around atoms along the X, Y, and Z axes, making it difficult to model their momentum space with a fixed graph structure.

To achieve the desired flexibility, Li and his collaborators introduced the concept of virtual nodes.

They created a model called the Virtual Node Graph Neural Network (VGNN), which represents phonons by adding a series of flexible virtual nodes to the fixed crystal structure. The virtual nodes allow the size of the neural network's output to vary, thus not being limited by the fixed crystal structure.

Virtual nodes are connected to the graph in a special way, only receiving information from real nodes. Although they are updated during the computation process as the model updates real nodes, they do not affect the accuracy of the model."Our approach is very efficient in terms of coding. You just need to generate a few extra nodes in your GNN. The physical location is not important, and the real nodes are not even aware of the existence of the virtual nodes," said Chotrattanapituk.

Reducing Complexity

Thanks to the virtual nodes representing phonons, VGNN can skip many complex calculations when estimating the phonon dispersion relationship, making this method more efficient than the standard GNN. The researchers proposed three versions of VGNN with different levels of complexity, each of which can predict the phonon dispersion relationship directly from the atomic coordinates of the material.

Their method has the flexibility to quickly simulate high-dimensional properties, which can be used to estimate the phonon dispersion relationship in alloy systems. For traditional methods, these complex combinations of metals and non-metals are particularly difficult to model.Researchers have also found that VGNN provides slightly higher accuracy in predicting the thermal capacity of materials. In some cases, the prediction error is two orders of magnitude lower than when using their technology.

Li stated that VGNN can calculate the phonon dispersion relations of thousands of materials within a few seconds on a personal computer.

This efficiency can enable scientists to search a larger space when looking for materials with specific thermal properties, such as excellent heat storage, energy conversion, or superconductivity.

Furthermore, the virtual node technology is not limited to phonons alone; it can also be used to predict other challenging properties such as optical and magnetic properties.

In the future, researchers hope to improve this technology to make virtual nodes more sensitive in order to capture subtle changes that may affect the phonon structure."Researchers have been too accustomed to using graph nodes to represent atoms in the past, but we should rethink this. Graph nodes can be anything. Virtual nodes are a very general method that can be used to predict a large number of high-dimensional quantities," said Li.

Olivier Delaire, an associate professor in the Department of Mechanical Engineering and Materials Science at Duke University, commented, "Although I did not participate in this work, the innovative method of the authors greatly enhances the description of solids by graph neural networks, incorporating key physical information elements such as wave vector-dependent band structures and dynamical matrices through virtual nodes. I find the acceleration in predicting complex phonon properties to be astonishing, several orders of magnitude faster than the most advanced general machine learning interatomic potentials. And it is impressive that this advanced neural network can capture fine features and follow physical laws. The model has great potential to be extended to describe other important material properties: electronic, optical, as well as magnetic spectra and band structures are directions that can be considered."

This work was supported by the U.S. Department of Energy, the National Science Foundation, the Mathworks Fellowship, the Sow-Hsin Chen Fellowship, the Harvard Quantum Initiative, and Oak Ridge National Laboratory.