The so-called “chemical space” of unknown materials and molecules could be said to be as large and unexplored as space itself. So there is good reason to develop methods for finding hitherto unknown materials. Researchers at Linköping University and the University of Cambridge have developed a machine learning method that can map chemical space on a greater scale than previously.
The AI model describes the symmetry between where atoms sit, which makes it easier to explore interesting possibilities.
“We have trained the model using over 300,000 materials, and have got it to suggest previously unknown materials in which the atoms are placed in new symmetrical ways”, says Rickard Armiento, docent in physics-based modelling at the unit for Materials Design and Informatics at Linköping University.
Less hay
Using the new method, new combinations of substances in new crystal structures can be predicted using AI, instead of needing to be manufactured in a lab. This speeds up the design and development of materials. This way, researchers can develop potential new materials for, for example, the development of batteries and solar cells.
“If we imagine materials discovery as searching for a needle in a haystack, our approach lets us dramatically reduce the amount of 'hay' before we begin our search”, says Rhys Goodall, PhD student at the Cavendish Laboratory at the University of Cambridge.
Searching for combinations of substances that can form stable materials is foundational for materials science, as well as for understanding how the material is structured. Determining whether a material is stable involves substantial computing. The model developed in this research makes computation much more effective, as the model studies existing materials to predict whether new combinations might be stable.
“We show how we can use this model to screen potential materials and focus our computational and experimental efforts on those that are most promising”, says Rhys Goodall.
Functional materials
The method predicts structures for promising materials that might, for example, be used for five times greater effectiveness in piezoelectricity and energy extraction.
Felix Andreas Faber is a postdoc at the University of Cambridge. Felix Andreas Faber, postdoc at the University of Cambridge. Photo credit Anders Törneholm He adds that current simulations that can calculate crystalline stabilities are very slow and costly.
“The space of theoretically possible inorganic solids is so big that it is impossible to investigate or make even a fraction of it. Our model overcomes many of these issues”, says Felix Andreas Faber.
The researchers now use the new model to search for new functional materials. The Division of Theoretical Physics at Linköping University has several projects the methods of which may some day be used for the development of material for hard surface coating, and for the design of defects for application within quantum information science.
The study was funded by the Swedish Research Council, Swedish e-Science Research Centre, as well as the Royal Society and Winton Programme for the Physics of Sustainability. The study has utilised computational resources from the National Supercomputer Centre at Linköping University, provided by the Swedish National Infrastructure for Computing.
The article: Rapid discovery of stable materials by coordinate-free coarse graining, Rhys A. Goodall, Abhijith S. Parackal, Felix A. Faber, Rickard Armiento, Alpha A. Lee, Science Advances published online 27 July 2022. DOI: 10.1126/sciadv.abn4117
Felix Andreas Faber, postdoc at the University of Cambridge, Rickard Armiento, associate professor at IFM and Abhijith S. Parackal, PhD student also at IFM, are three of the authors behind the paper published in Science Advances. Photo credit Anders Törneholm
