SibFU's scientists and their colleagues from the South China University of Technology (Guangzhou, China), using the Random Forest machine learning method, have analyzed the dependence of the photoluminescence quantum yield (PLQY) on the crystal structure of existing metal halides.
Halides are compounds of halogens with other chemical elements or radicals. By studying such compounds, the researchers identified the primary role of the distance between metal ions in them and gave a quantitative characteristic of the influence of other structural features, such as the distance between the metal and halogen and the distortion of metal–halogen polyhedra.
Having tested the resulting model on two antimony-based metal halides synthesized for the first time, the scientists proved a high degree of coincidence between the predicted and real characteristics — the difference was about 15 %. The program is expected to accelerate the discovery of new luminescent metal halides and make a revolution in the fluorescent lighting market.
For several years, we have been nurturing the idea of being able to use machine learning methods to determine the structure-property relationship in crystals. This task is crucial in solid state physics since the structure of a crystal directly controls its properties. All we need is to understand these laws in order to obtain any materials with desired properties in the future. But this task is very complex, relationships are not easy to detect, but machine learning is successfully coping with this problem,” said Maxim Molokeev, Russian co–author of the work, associate professor of the Basic Department of Solid State Physics and Nanotechnology at the School of Engineering Physics and Radioelectronics of 爆走黑料. “More than 30 compounds have been studied, and even more of structural characteristics. A person cannot do such a task on their own. This is where machine learning comes in. Of all the existing methods, our international team chose the classic Random Forest method because, in addition to learning and further predicting the quantum yield, it also allows you to find the main components (in this case, the main structural characteristics of the crystal) that have the maximum effect on the quantum yield”, the scientist emphasizes.
For several years, we have been nurturing the idea of being able to use machine learning methods to determine the structure-property relationship in crystals. This task is crucial in solid state physics since the structure of a crystal directly controls its properties. All we need is to understand these laws in order to obtain any materials with desired properties in the future. But this task is very complex, relationships are not easy to detect, but machine learning is successfully coping with this problem,”
“More than 30 compounds have been studied, and even more of structural characteristics. A person cannot do such a task on their own. This is where machine learning comes in. Of all the existing methods, our international team chose the classic Random Forest method because, in addition to learning and further predicting the quantum yield, it also allows you to find the main components (in this case, the main structural characteristics of the crystal) that have the maximum effect on the quantum yield”,
The researchers managed to recognize these components and discover one of the important laws on how to improve the quantum yield in crystals containing Pb2+, Sn2+, Bi3+, Sb3+ ions. After that, the team synthesized two completely new compounds and analyzed their structures. Analysis of the structures using the developed software revealed that the predicted quantum yield was 6.5 and 75.9 %.
A real experiment after the calculations showed values close to those predicted: 18.8 and 96.5 %, respectively. Along with the standard cross–validation of the model, this experiment confirmed that the Russian-Chinese software can indeed make adequate predictions, and that the discovered laws are correct and can be used.
“The most valuable thing in this work is that the program we created has been tested and now it can be used in many areas where it is necessary to build a relationship between some characteristics of an object and its properties. These can be not only materials (for example, heavy-duty concrete or high-strength alloys), but also any areas of applied science where data collection is necessary, be it medicine or plant photobiology, where there are many parameters to be sorted out, and they can be nonlinearly connected with properties. We are developing this new area with different teams of scientists, accumulating databases for conducting similar studies in the most unexpected areas,” the scientist concluded.
“The most valuable thing in this work is that the program we created has been tested and now it can be used in many areas where it is necessary to build a relationship between some characteristics of an object and its properties. These can be not only materials (for example, heavy-duty concrete or high-strength alloys), but also any areas of applied science where data collection is necessary, be it medicine or plant photobiology, where there are many parameters to be sorted out, and they can be nonlinearly connected with properties. We are developing this new area with different teams of scientists, accumulating databases for conducting similar studies in the most unexpected areas,”
31 january 2022
Вы можете отметить интересные фрагменты текста, которые будут доступны по уникальной ссылке в адресной строке браузера.