An international team of researchers with the participation of young scientists from the HSE Faculty of Computer Science and Artificial Intelligence Centre have developed a machine learning algorithm that can determine the properties of new 2D materials with point defects. The new method is 1000 times faster than quantum mechanical computations and 3.7 times more accurate than other machine learning algorithms. The results have been published in npj Computational Materials. The source code, dataset, and model weights are available in the repository under an open licence.
Two-dimensional materials are currently experiencing a boom in interest. These are exceptionally thin crystal lattices, with a thickness of only one or a few atoms. When a crystal layer becomes as thin as this, the bonds between its compounds change, causing the material to exhibit unique electrochemical properties.
The most famous example is graphene, which consists of a single layer of carbon atoms. Much has been written about the unique properties of graphene: it is remarkably durable, flexible, transparent, exhibits high electrical and thermal conductivity, and boasts record mobility of charge carriers. The combination of all these properties makes graphene, along with other 2D materials, ideal candidates for high-tech solutions. Thus, 2D materials are already widely used in the production of transistors, sensors, biosensors, solar panels, ultrathin lenses, highly sensitive screens, and various other applications.
In recent years, scientists have successfully synthesised many types of 2D materials; however, their properties are still poorly understood. This is attributed to a distinctive feature of 2D materials – the presence of defects in their structure, which can significantly influence their characteristics.
Such defects can occur randomly during the production process, as it is difficult to produce large sheets of graphene and similar materials without breaks or folds. Alternatively scientists might intentionally want to replace one atom with another or remove an atom from the compound to observe the resulting changes in the material. Thus, a specific combination of defects can help achieve the desired properties and create a new material, e.g., one that can be used in highly efficient solar panels.
Researchers from the HSE Faculty of Computer Science and Artificial Intelligence Centre, in collaboration with their colleagues from Innopolis University, the National University of Singapore, and Constructor University Bremen (Germany), have developed a machine learning algorithm that enables rapid and accurate prediction of the properties of 2D materials by analysing their defects. The main difference from other existing models is that this new algorithm allows scientists to work with multiple defects simultaneously and analyse their configuration within the material. In contrast, other well-known methods, which are time-consuming and expensive, involve a sequential analysis of each individual defect.
Our algorithm can simultaneously handle multiple defects. Additionally, it exclusively analyses defects rather than all atoms in the structure, as other models do. That is why it operates 1000 times faster than quantum mechanical computation and achieves 3.7 times greater accuracy compared to its contenders.
In recent years, thanks to a large number of studies,comprehensive databases of 2D materials have been compiled, making it possible to apply ML methods to various problems in this field. Our algorithm has been pre-trained on a vast number of samples, enabling us to precisely determine how a specific configuration of defects will impact the material's properties.
With the help of our model, we can answer the question of what kind of material needs to be produced for a specific task. For example, we could intuitively hypothesize that by taking boron nitride and replacing 3% of its atoms with carbon atoms, while 'knocking out' 2% of the atoms, we can produce an efficient solar panel. Our algorithm enables us to rapidly test this hypothesis. While we cannot guarantee that the idea will necessarily work, we can eliminate the many options that definitely won't work.
The proposed model represents a significant advancement towards controllable defect engineering in 2D materials. While neural networks may not be capable of generating something fundamentally new, since they are trained using existing data, they excel at rapidly and efficiently processing and summarising information. According to the authors, the next stage of their research will involve developing a model that, based on the analysis of materials and potential configurations of defects, will propose potential new candidates for specific high-tech solutions.