Machine Learning Potential Aided Structure Search For Au Clusters

The investigation of the atomic structures of materials is crucial for the understanding of their mechanical,electrical,magnetic,and thermal properties.Many crystal structure prediction methods have been developed in the past,including the traditional structure searching methods,such as genetic algorithm,and machine-learning-based searching methods.On the other hand,clusters have attracted substantial research interests due to their unique physical and chemical properties,such as excellent catalytic activity of Au clusters,and have promised great potential in the application of nano-electronics,nano-biology,and nano-catalysis.In this thesis,we develop a computational framework based on machine-learning potentials for global structure optimization and apply it to determine the stable structures of a series of Au clusters as a demonstration.The first chapter introduces the research status of structure prediction methods and machine learning algorithms.Firstly,the basic steps of structure prediction methods are introduced,and then the working scenarios and basic processes of machine learning are described.The second chapter introduces our computational framework for global structure optimization.We first introduce Bayesian optimization and BPNN(Behler-Parrinello Neural Network)algorithms.For deep learning algorithms,we introduce graph neural network and SchNet.Finally,active learning,which seeks to obtain the best performance with the least training data,and particle-swarm global optimization algorithm are introduced.The third chapter presents our results of using the developed SchNet potentials aided particle-swarm optimization method to search the most stable structures of Au clusters.Firstly,SchNet potentials are obtained by training on datasets containing Au clusters with different sizes.Using these potentials,we identify the most stable structures of a series of Au clusters in the sizes from 4 to 20 atoms.We find that small Au clusters less than 14 atoms tend to form two-dimensional planar structures,while large ones tend to be three-dimensional structures.In addition,we show the performance of Bayesian optimization and active learning on the structure searching of Au20 and Au13 clusters,respectively.The fourth chapter summarizes the research work of this thesis and provide an outlook to this field.