Research On New Nano-cluster Structure Prediction Based On Particle Swarm And Gaussian Process Regression

Structure is the key to the properties of materials.To deeply understand the properties of matter we must clarify the structure of it.Because of the influences of various external factors,it is quite challenging to directly determine the structure of materials through experimental approaches,which will make researches inefficient.Using computer simulation and calculation to obtain the structures of substances will not be affected by environmental factors,and can provide guidance for manufacturing the target material.In recent decades,many methods have been proposed and applied to structural prediction.But these methods still have some shortcomings.There are two main difficulties in material structure prediction:1.Existing optimization algorithms have low computational efficiency and success rate in the prediction of large-scale systems;2.The methods based on density functional theory is very accurate but the calculation is too time-consuming.In this paper,taking nanoclusters as the research object,we have made very in-depth study on these two major issues.The main contents of this paper are:Based on particle swarm optimization algorithm,combined with random learning,competition mechanism and mutation operator,an efficient and unbiased optimization algorithm is proposed.For clusters,initial solution generation based on point group symmetry restriction and similar structure exclusion based on bond characterization matrix are proposed.In addition,a cluster structure optimization calculation software package with high usability was developed.Furthermore,some detailed structural predictions were performed on a variety of nanoclusters with different sizes and different chemical compositions,and the algorithm achieved good convergence performance.In the calculation based on density functional theory,the algorithm discovered new global optimal structures of Pt9?Pt11?Pt12?Pt15?Ag16 and Ag17,which proved the predictive ability of the algorithm.Aiming at the problem that density functional theory calculation is very time-consuming,an accelerated structural prediction method is proposed based on global optimization and machine learning.In structure prediction of Au20 cluster,using this machine learning accelerated method can save nearly 3000 hours of calculation,which greatly improved the efficiency of structure prediction.The result shows that this acceleration method can effectively reduce the calculation time required in structural prediction and make the prediction of large size material structures possible,which could mark up as a breakthrough in the prediction of material structure.