Reviewed by Danielle Ellis, B.Sc.Oct 8 2024
A new artificial intelligence (AI) tool for producing high-quality optical spectra with the same accuracy as quantum simulations but operating a million times faster has been revealed by researchers from Tohoku University and the Massachusetts Institute of Technology (MIT) in a study published in Advanced Materials. This tool could hasten the development of quantum and photovoltaic materials.
An AI tool called GNNOpt can accurately predict optical spectra based solely on crystal structures and speed up the development of photovoltaic and quantum materials. Image Credit: Nguyen Tuan Hung et al.
Developing optoelectronic devices, including LEDs, solar cells, photodetectors, and photonic integrated circuits, requires an understanding of the optical characteristics of materials. These devices are essential to the current resurgence of the semiconductor industry.
It is challenging to rapidly test a large number of materials using traditional methods of calculation based on the fundamental laws of physics because they require intricate mathematical computations and a significant amount of processing power.
By overcoming this obstacle, new photovoltaic materials for energy conversion may be discovered, and the optical spectra of materials may provide a deeper understanding of their underlying physics.
This is exactly what a team led by Mingda Li, an associate professor at MIT's Department of Nuclear Science and Engineering (NSE), and Nguyen Tuan Hung, an assistant professor at the Frontier Institute for Interdisciplinary Science (FRIS), Tohoku University, accomplished by introducing a new AI model that uses only a material’s crystal structure as an input to predict optical properties across a wide range of light frequencies.
Nguyen, the lead author, and his colleagues recently published their findings in Advanced Materials, an open-access journal.
Optics is a fascinating aspect of condensed matter physics, governed by the causal relationship known as the Kramers-Krönig (KK) relation. 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.
Nguyen Tuan Hung, Study Lead Author and Assistant Professor, Frontier Institute for Interdisciplinary Science (FRIS), Tohoku University
The limitations of laser wavelengths make it difficult to obtain optical spectra with complete frequency coverage in experiments. Simulations are also complex, necessitating stringent convergence criteria and incurring substantial computational costs. As a result, the scientific community has long sought more effective methods for predicting the optical spectra of various materials.
Machine-learning models utilized for optical prediction are called graph neural networks (GNNs). GNNs provide a natural representation of molecules and materials by representing atoms as graph nodes and interatomic bonds as graph edges.
Ryotaro Okabe, Graduate Student, Massachusetts Institute of Technology
However, while GNNs have shown promise in predicting material properties, they are not universally applicable, particularly in representing crystal structures. To solve this conundrum, Nguyen and colleagues developed a universal ensemble embedding, in which multiple models or algorithms are created to unify the data representation.
This ensemble embedding goes beyond human intuition but is broadly applicable to improve prediction accuracy without affecting neural network structures.
Abhijatmedhi Chotrattanapituk, Graduate Student, Massachusetts Institute of Technology
The ensemble embedding method is a universal layer that can be applied to any neural network model without changing its structure.
“This implies that universal embedding can readily be integrated into any machine learning architecture, potentially making a profound impact on data science,” added Mingda Li.
This method allows for highly accurate optical prediction based solely on crystal structures, making it suitable for a wide range of applications, including screening materials for high-performance solar cells and detecting quantum materials.
In the future, the researchers hope to improve the AI model’s ability to predict material properties based only on crystal structures by creating new databases for characteristics like mechanical and magnetic properties.
Journal Reference:
Hung, N. T. et. al. (2024) Universal Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structures. Advanced Materials. doi.org/10.1002/adma.202409175