Machine Learning to Predict the Electronic Band Structure of Materials
AI and machine learning have real potential in materials science. Researchers have been exploring these techniques to predict the electronic band structure of materials directly from band mapping data.
Why band structure matters
The electronic band structure of a material is a fundamental property determining its electrical and optical behaviour. It describes the range of energy levels electrons in the material can occupy and how those levels are distributed.
Traditionally band structure is determined experimentally, through methods such as angle-resolved photoemission spectroscopy. Those experiments are time-consuming and expensive, and are not feasible for every material.
The machine learning approach
To get around that, researchers train algorithms on large datasets of materials and their corresponding band structures, then predict the band structure of a new material from its band mapping data.
The proposed method uses a probabilistic machine learning model fitted to the band mapping data, aiming to predict the energy values of the material's band structure accurately. The model uses a nearest-neighbour Gaussian distribution to describe the proximity of energy values at nearby momenta, and finds the optimal fit through maximum a posteriori estimation in probabilistic inference.
The distinctive part is that it can incorporate imperfect physical knowledge, such as the presence of impurities or defects, and handle noise in the data. That makes the model robust in real-world conditions where the data is never clean.
Two families of method
Machine learning based band structure prediction trains on a dataset of materials and their band structures, then predicts for a new material from its band mapping data.
Deep learning based band structure prediction does the same using a deep neural network, which learns the complex relationships between band mapping data and the resulting band structure.
Why it matters
Using these techniques to predict band structure from mapping data could substantially accelerate the discovery and development of new materials. It may also open up the study of materials that are difficult or impossible to examine with traditional experimental methods.