TU-TESL

• 22

(picture source: Argonne official website) it is a complex work to design the best molecular composition for components. It's like creating new cake recipes based on billions of potential ingredients. Designers need to face many challenges, such as determining which ingredients are most appropriate to match. Moreover, even with state-of-the-art supercomputers, scientists cannot accurately simulate the chemical properties of each molecule to prove that it can become the basis for the next generation of battery materials. According to foreign media reports, researchers at the Argonne National Laboratory of the U.S. Department of energy have greatly accelerated the battery research and development process with the help of machine learning and artificial intelligence. First, the researchers use g4mp2 to calculate the intensive model and build a highly accurate database, which contains about 133000 small organic molecules, which may constitute the basic electrolyte of the battery. However, these are just a small part of the 166 billion molecules that scientists want to study. In order to save computing time and power, the research team uses machine learning algorithm to link the precise known structure in small data groups with the coarser modeling structure in large data groups. "We believe that machine learning represents a method," said Ian foster, director of Argonne's data science and learning department. It takes only a small part of the calculation cost to obtain a nearly accurate molecular image. " In order to lay a foundation for machine learning model, foster and his colleagues, based on the density functional theory, use the modeling framework with less computation, and use the quantum mechanics modeling framework to calculate the electronic structure of large-scale system. Density functional theory can explain molecular properties better, but it is not as accurate as g4mp2. In order to better and more widely understand the information of organic molecules, it is necessary to use g4mp2 with high precision to calculate the positions of atoms in molecules, and compare it with the molecules analyzed only by density functional theory to improve the algorithm. Using g4mp2 as the standard, the researchers trained the DFT model and added correction factors to reduce the calculation cost and improve the accuracy. "Machine learning algorithms provide us with a way to study the relationship between atoms and their neighbors in macromolecules, understand how they combine and interact, and look for similarities between these molecules and other familiar molecules," said Logan ward, a computing scientist at Argonne. On this basis, we can predict the energy of macromolecules, or the difference between high and low accuracy calculations. " "We are launching this project to get the largest possible image of the candidate components of the battery electrolyte," said Rajeev assary, a Argonne chemist. To use a molecule for energy storage applications, we need to understand its properties, such as stability. Through machine learning, we can predict the properties of macromolecules more accurately. "