image: This is a neural network protocol for predicting molecular bond energy.
Theoretical prediction of molecular bond energy is of key importance for understanding molecular properties. For instance, information of chemical reaction pathways can be inferred from a numerical analysis of the bond energies involved. A well-accepted way is to calculate the difference in values of molecular energies before and after chemical bond rupture using quantum chemistry tools. However, for complicated chemical reactions, a large number of chemical bonds need to be analyzed, requiring many repeated quantum mechanics calculations that consume heavy computing resources. Other ways such as mapping method are unreliable and limited. Developing an efficient solution for quickly and accurately predicting bond energies remain an open challenge.
Recently, data-driven research paradigm based on big data and artificial intelligence (AI) techniques is increasingly important in chemistry. Especially, neural networks ---- a class of algorithms featured by learning characters of data, is deemed a promising AI approach that can significantly reduce the computation cost of complex problems with well-defined rules yet high-dimensional data. "Inspired by this," Prof. Jiang Jun introduced, a research group leader at Hefei National Laboratory for Physical Science at the Microscale of University of Science and Technology of China, "we have employed neural networks to predict the molecular bond energies. In addition, we found that the combination of artificial intelligence and theoretical calculations of quantum chemistry provides an efficient tool for accurately and quickly predicting molecular bond energy."
"Recent years, our group has devoted ourselves to the development of the application of machine learning technology in the field of quantification, and has tried to make it an important tool for solving quantitative problems", Jiang said, "in this work, we first obtained 8,000 sets of bond energy data based on quantum chemical calculations. Through the random forest method, the appropriate descriptors are selected from the basic information (e.g. geometry and charge) of molecules, and the 8,000 sets of quantum chemical data are iteratively learned through the neural network, and the neural network model between the molecular bond energy and its basic state information is established. This model successfully predicted the molecular bond energy."
In addition, the performance of neural network models with different kinds of atomic charge distributions is also compared. This work proves the feasibility and advantages of machine learning in molecular bond energy simulation, and provides a reasonable solution for predicting molecular bond energy in complex systems.