Photo credit: phys.org
Revolutionizing Material Discovery: A New AI Approach to Calculate Electron Structures
The process of determining a material’s electron structure can demand extensive computational resources, often taking up to a million CPU hours. However, a breakthrough from a team of researchers at Yale University aims to accelerate these calculations through advanced artificial intelligence techniques, enabling quicker and more accurate material discovery. Their findings have recently been detailed in a publication in Nature Communications.
The Challenge of Electronic Structure Analysis
Understanding the electronic structure of materials is critical in materials science, as it provides insights into complex systems, such as moiré systems and defect states. The traditional approach utilizes density functional theory (DFT) to analyze electronic structures, which functions well for many scenarios. However, Professor Diana Qiu, who led the study, emphasizes its limitations in certain contexts. “When examining excited state properties—like a material’s interaction with light or its electrical conductivity—DFT falls short in delivering comprehensive insights,” she explained.
To address these limitations, researchers often resort to more sophisticated theories that extend beyond DFT. Yet, Qiu’s research grapples with computational cost, making it challenging to apply these higher-level theories effectively. While machine learning techniques have been explored, they do not adequately resolve the complexity inherent in determining material band structures, which are essential for comprehending material properties.
Innovative Use of Wave Functions
In response to these challenges, Qiu and her team decided to delve into the electrons’ wave functions—mathematical representations of a particle’s quantum state. Their study specifically focused on two-dimensional (2D) materials, which consist of only a few atomic layers.
“Our investigation aimed to decipher the wave function,” Qiu stated. “This probability distribution can be visualized as an image in space.”
Leveraging Variational Autoencoders
The researchers employed a variational autoencoder (VAE), an artificial intelligence tool used for image processing, to create a compact representation of the wave function.
“This process is executed in an unsupervised manner, meaning there’s no human oversight involved,” Qiu explained. “It compresses a 100-gigabyte dataset into 30 numbers that encapsulate the initial wave function. These numbers then act as inputs for a second neural network designed to predict more elaborate excited state properties.”
This method not only provides a more precise representation that is less influenced by human intuition but also enhances efficiency and applicability across various areas of research.
“Our resultant representation is about 100 to 1,000 times smaller than what is typically necessary for feature selection,” Qiu noted, highlighting a critical aspect—simplifying input enhances versatility in application.
Significant Time Savings
Traditionally, calculating the band structure of a three-atom material could require between 100,000 and a million CPU hours. In stark contrast, the VAE-assisted approach developed by Qiu’s team reduces this time to merely an hour.
“The primary practical advantage of this method is its ability to significantly expedite complex calculations,” she remarked. “This expediency allows us to examine a broader range of materials, facilitating the discovery of new materials with desirable properties.”
More information: Bowen Hou et al, Unsupervised representation learning of Kohn–Sham states and consequences for downstream predictions of many-body effects, Nature Communications (2024). DOI: 10.1038/s41467-024-53748-7
Citation: Faster way to calculate electron structure makes it easier to discover new materials (2024, December 19) retrieved 19 December 2024 from https://phys.org/news/2024-12-faster-electron-easier-materials.html
This research marks a significant advancement in the toolkit available to materials scientists, potentially altering the landscape of material discovery and development in the years to come.
Source
phys.org