Using artificial intelligence, a study of 5.5 million compounds identified hundreds of surfaces capable of holding on to their electrons either very strongly or very weakly. This could prove valuable for the electronics and batteries of the future.
To understand the significance of this result, we need to look at what happens at the surface of a material. An electron cannot always leave it easily. A certain amount of energy must be supplied. This depends in particular on the material and on how its atoms are arranged at the surface.
A low value makes it easier for electrons to escape. A high value holds on to them more strongly. This property plays a role in many electronic components, in certain energy-conversion systems, and in chemical reactions involving surfaces. Finding materials at the extremes could therefore open up possibilities for designing new devices.
Testing millions of candidates with the most precise computational methods would nevertheless require enormous computing resources. Jun Meng and his colleagues therefore proceeded in stages. Their system starts with a fast machine-learning AI model, then reserves the costly physical calculations for candidates that deserve more thorough verification.
The model also takes its level of uncertainty into account. In practice, a prediction deemed unreliable can be checked with a more precise method instead of being accepted as is. The researchers applied this screening process to 5.5 million compounds from the GNoME and Alexandria databases.
The screening identified 209 surfaces with a work function below 2.0 electronvolts. They correspond to 136 distinct materials. At the other extreme, 227 surfaces exceed 6.0 electronvolts, representing 172 materials.
Some results follow already known chemical trends. For example, surfaces terminated with alkali or alkaline-earth metals often appear among those that release their electrons more easily. More interestingly, the analysis also identified less familiar combinations, including lanthanide-rich surfaces among those with particularly low values.
The next step is to examine the most promising candidates experimentally. Their stability, manufacturability, and real-world behavior will need to be verified. Above all, the researchers’ work provides a method for rapidly narrowing down a list of several million possibilities before devoting experiments and intensive computations to a few materials.