Researchers from Tohoku University and the Massachusetts Institute of Technology (MIT) have introduced an advanced AI tool capable of generating high-quality optical spectra with an accuracy comparable to that of quantum simulations, while operating at a speed that is a million times faster. This innovation has the potential to significantly accelerate the development of photovoltaic and quantum materials.
The ability to accurately understand the optical properties of materials is crucial for the development of optoelectronic devices, such as light-emitting diodes (LEDs), solar cells, photodetectors, and photonic integrated circuits. These devices play a critical role in the current resurgence of the semiconductor industry.
Traditional methods for calculating optical properties rely on fundamental physical laws and involve complex mathematical computations, requiring significant computational resources. This makes it difficult to rapidly test a large number of materials. Overcoming this limitation could facilitate the discovery of new photovoltaic materials for energy conversion, as well as deepen our understanding of the fundamental physics underlying materials’ optical spectra.
A research team led by Nguyen Tuan Hung, an Assistant Professor at the Frontier Institute for Interdisciplinary Science (FRIS) at Tohoku University, and Mingda Li, an Associate Professor at MIT’s Department of Nuclear Science and Engineering (NSE), has addressed this challenge by developing a novel AI model. This model is capable of predicting optical properties across a broad range of light frequencies, requiring only the material’s crystal structure as input.
“Optics is a fascinating aspect of condensed matter physics, governed by the causal relationship known as the Kramers-Krönig (KK) relation,” says Nguyen. “Once one optical property is known, all other optical properties can be derived using the KK relation. It is intriguing to observe how AI models can grasp physics concepts through this relation.”
Obtaining optical spectra with comprehensive frequency coverage presents a significant challenge in experimental settings due to the limitations in the range of available laser wavelengths. Simulations, while providing valuable insights, are also highly complex, often requiring strict convergence criteria and substantial computational resources. These limitations have prompted ongoing efforts within the scientific community to develop more efficient methods for predicting the optical spectra of diverse materials. The pursuit of such methods is critical for advancing our understanding of material properties and for facilitating the development of technologies in areas like photovoltaics and optoelectronics.
“Machine-learning models utilized for optical prediction are called graph neural networks (GNNs),” points out Ryotaro Okabe, a chemistry graduate student at MIT. “GNNs provide a natural representation of molecules and materials by representing atoms as graph nodes and interatomic bonds as graph edges.”
While Graph Neural Networks (GNNs) have demonstrated potential in predicting material properties, they lack universality, particularly in accurately representing complex crystal structures. To address this limitation, Nguyen and colleagues developed a universal ensemble embedding approach. This method involves creating multiple models or algorithms that work in tandem to unify and standardize the representation of data, providing a more robust and comprehensive framework for predicting the optical properties of materials across diverse structures. This innovation enhances the generalizability of AI models in material science, overcoming the limitations of traditional GNNs.
“This ensemble embedding goes beyond human intuition but is broadly applicable to improve prediction accuracy without affecting neural network structures,” explains Abhijatmedhi Chotrattanapituk, an electrical engineering and computer science graduate student at MIT.
The ensemble embedding method functions as a universal layer that can be effortlessly incorporated into any neural network model without necessitating modifications to the network’s architecture. According to Mingda Li, “This implies that universal embedding can readily be integrated into any machine learning architecture, potentially making a profound impact on data science.”
This approach enables highly accurate optical property predictions using only the crystal structure of a material, making it highly applicable for tasks such as screening materials for high-performance solar cells and identifying quantum materials.
- See also: Toward a code-breaking quantum computer
In the future, the researchers plan to extend their work by developing new databases that encompass various material properties, including mechanical and magnetic characteristics. This would further enhance the AI model’s ability to predict a broader range of material properties solely from crystal structure information, expanding its potential applications across different fields in materials science.
Lead author Nguyen and his colleagues published their findings in Advanced Materials.
