A new machine learning model for characterizing material surfaces

thank you. Listen to this article using the player above. ✖

The design and development of new materials with superior properties requires comprehensive analysis of their atomic and electronic structures. Electron factors such as ionization potential (IP), which is the energy required to remove an electron from the valence band maximum, and electron affinity (EA), which is the amount of energy released when an electron binds to the conduction band minimum. Energy parameters reveal important information. Information about the electronic band structure of the surfaces of semiconductors, insulators, and dielectrics. Accurate estimation of the IP and EA of such nonmetallic materials can demonstrate their applicability for use as functional surfaces and interfaces in photosensitive and optoelectronic devices.

Furthermore, IP and EA are highly dependent on the surface structure, which adds another dimension to the complex procedure of their quantification. Traditional calculations of IP and EA involve the use of exact ab initio calculations, where the bulk and surface systems are quantified separately. This time-consuming process makes it impossible to quantify the IP and EA of many surfaces and requires the use of computationally efficient approaches.

Professor Ohba said, “ML has attracted a lot of attention in materials science research in recent years.The ability to virtually screen materials based on ML technology is an extremely efficient way to search for new materials with excellent properties.'' “This is a practical method,'' explains the motivation behind this research. Additionally, large datasets can be trained using accurate theoretical calculations, allowing for better prediction of important surface properties and their functional impact. ”

The researchers used an artificial neural network to develop a regression model that incorporated smooth overlap of atomic positions (SOAP) as numerical input data. Their model accurately and efficiently predicted the IP and EA of binary oxide surfaces using information about the bulk crystal structure and surface termination planes.

Additionally, ML-based predictive models are subject to “transfer learning,” a scenario in which a model developed for a specific purpose can be created, incorporate new datasets, and be reapplied to additional tasks. Scientists incorporated the effects of multiple cations into the model by developing a “learnable” SOAP and used transfer learning to predict the IP and EA of ternary oxides.

Professor Ohba concludes, “Our model is not limited to predicting surface properties of oxides, but can be extended to study other compounds and their properties.”

reference: Kiyohara S, Hinuma Y, Oba F. Band alignment of oxides using learnable structural descriptor-assisted neural networks and transfer learning. J Am Kem Soc. 2024;146(14):9697-9708. doi: 10.1021/jacs.3c13574

This article has been reprinted from the following material: Note: Materials may be edited for length and content. Please contact the citation source for details.You can access our press release publishing policy here.