Finding new LEDs, solar cells, and photodetectors requires a deep understanding of materials’ optical properties, which is usually a time-consuming and resource-heavy process. However, researchers have introduced a groundbreaking AI tool that predicts these optical properties much faster than traditional quantum simulations.
A collaborative team from Tohoku University and the Massachusetts Institute of Technology (MIT) has developed a new AI tool capable of generating high-quality optical spectra with the same level of precision as quantum simulations, but at a speed that is a million times quicker. This innovation could greatly enhance the development of photovoltaic and quantum materials.
Grasping the optical traits of materials is crucial for creating optoelectronic devices, including LEDs, solar cells, photodetectors, and photonic integrated circuits. These devices play a key role in the current revival of the semiconductor industry.
Conventional methods for calculating optical properties involve intricate mathematical equations and substantial computational resources, making it challenging to quickly evaluate a broad range of materials. Overcoming this barrier could lead to the identification of novel photovoltaic materials for energy conversion, as well as a richer understanding of the basic physics of materials through their optical spectra.
A team led by Nguyen Tuan Hung, an assistant professor at Tohoku University’s Frontier Institute for Interdisciplinary Science (FRIS), alongside Mingda Li, an associate professor in MIT’s Nuclear Science and Engineering Department, achieved this by creating a new AI model that can forecast optical properties over a wide spectrum of light frequencies by utilizing only a material’s crystal structure as input.
The lead author, Nguyen, and his colleagues shared their research findings in an open-access paper published in Advanced Materials.
“Optics is a fascinating domain within condensed matter physics, regulated by the Kramers-Krönig (KK) relation,” Nguyen explains. “Once one optical property is determined, all other characteristics can be derived using the KK relation. It is captivating to see how AI models can learn physics concepts through this relationship.”
Obtaining optical spectra that cover the full frequency range during experiments poses challenges due to the limitations of laser wavelengths. Simulations are also complicated, requiring strict convergence criteria and substantial computational expenses. Consequently, the scientific community has long sought more efficient ways to predict the optical spectra of different materials.
“The machine-learning models used for optical predictions are termed graph neural networks (GNNs),” notes Ryotaro Okabe, a chemistry graduate student at MIT. “GNNs effectively represent molecules and materials by modeling atoms as graph nodes and interatomic connections as graph edges.”
Despite the potential of GNNs to predict material properties, they are often not universally applicable, particularly regarding the representation of crystal structures. To address this issue, Nguyen and his team established a universal ensemble embedding that combines multiple models or algorithms to standardize data representation.
“This ensemble embedding extends beyond human intuition and is broadly applicable, enhancing prediction accuracy without altering the structures of the neural networks,” says Abhijatmedhi Chotrattanapituk, an electrical engineering and computer science graduate student at MIT.
The ensemble embedding technique serves as a universal layer that can be easily integrated into any neural network model without changing the underlying network design. “This means universal embedding can be readily incorporated into any machine learning framework, which could significantly influence data science,” remarks Mingda Li.
Thanks to this approach, the AI tool can deliver highly accurate optical predictions using only crystal structures, making it ideal for numerous applications, such as evaluating materials for advanced solar cells and identifying quantum materials.
Looking ahead, the researchers plan to create new databases containing various material properties, including mechanical and magnetic features, to further enhance the AI model’s ability to predict material properties based solely on crystal structures.
