Researchers from Princeton University created an instrument based on artificial intelligence (AI), which is able to predict the behavior of crystalline materials. The discovery is of great importance for the development of technology, including batteries and semiconductors. The new method uses a large language model similar to those used in text generators, for example, ChatGPT.
A distinctive feature of this approach is to synthesize information from text descriptions that contain details about the length and angles of the connections between atoms, as well as measure electronic and optical properties. Thus, the method can predict the properties of new materials more accurately and comprehensively than existing simulations, and potentially speed up the design and testing process of new technologies.
For teaching the adapted version of the T5 model, the initially created Google, the researchers have developed a reference text containing descriptions of more than 140,000 crystals from base data Materials Project. The effectiveness of the tool was checked by the example of the prediction of the properties of already studied crystalline structures, from ordinary table salt to silicon semiconductors. Now, after demonstrating the predictive force of T5, researchers are working on the use of a tool to design new crystalline materials.
The method was presented on November 29 at the Autumn Assembly of the Society for the Study of Materials ( materials Research Society ) in Boston. According to the authors of the work, the developed method is a new standard that can accelerate the opening of materials for a wide range of applications.
previously existing AI tools to predict the properties of crystals were based on the so -called graph neural networks, but they have limited computing power and cannot adequately capture the nuances of the geometry and the length of the connections between the atoms in the crystal. The team of scientists Princeton University was the first to solve this problem with the help of large language models.
However, as a new tool, the prediction method has its restrictions. It requires more computing power and works slower than graphic neural networks, usually used for this purpose. Experts strive for cooperation with other scholars of materials and plan to expand their work outside the crystals, covering a wider range of materials to improve the ability to predict the properties of new materials