Machine learning (ML) has been demonstrated to accurately and efficiently compute fundamental electronic properties of binary and ternary oxide surfaces, allowing for the screening of surface properties of materials and the development of functional materials. These properties, such as ionization potential (IP) and electron affinity (EA), provide crucial information about the electronic band structure of semiconductors, insulators, and dielectrics. The accurate estimation of IPs and EAs is essential for determining the potential applications of these materials in optoelectronic devices and photosensitive equipment.
Surface structures play a significant role in determining the IPs and EAs of nonmetallic materials, making their quantification a complex process. Traditional methods rely on time-consuming first-principles calculations, which often limit the ability to quantify these properties for many surfaces. To address these challenges, a team of scientists from Tokyo Tech led by Professor Fumiyasu Oba turned to machine learning as a more efficient approach. Their research, published in the Journal of the American Chemical Society, focuses on the development of a regression model using artificial neural networks and smooth overlap of atom positions (SOAPs) as input data for predicting IPs and EAs of binary oxide surfaces.
The ML-based prediction model developed by the researchers not only accurately predicted IPs and EAs of binary oxide surfaces but also demonstrated transfer learning capabilities. By incorporating learnable SOAPs and considering the effects of multiple cations, the model successfully predicted the IPs and EAs of ternary oxides using transfer learning. This versatility in the model’s application highlights its potential for studying a wide range of compounds and properties beyond oxide surfaces. The ability to train large datasets using accurate theoretical calculations enables the successful prediction of important surface characteristics and their functional implications.
Prof. Oba emphasizes the broader implications of their research, stating that ML technology allows for the efficient exploration of novel materials with superior properties by virtually screening materials based on large datasets. The model developed by the Tokyo Tech team could be extended to other compounds and properties, indicating its potential for advancing materials science research. By incorporating information on bulk crystal structures and surface termination planes, the ML-based regression model offers a more efficient and accurate method for predicting electronic properties of nonmetallic materials.
In conclusion, the research conducted by the team at Tokyo Tech demonstrates the effectiveness of machine learning in predicting surface properties of oxides and presents a valuable tool for the design and development of functional materials. The ability to accurately estimate IPs and EAs of binary and ternary oxide surfaces using ML-based models opens up opportunities for studying a wide range of compounds and properties. With its transfer learning capabilities and potential for extension beyond oxide surfaces, the model developed by the researchers shows promise for advancing materials science research and accelerating the discovery of novel materials with superior properties.
