Optimizating First Principle Software VASP Based On The New Sunway Supercomputer

With the development of Chinese super computer industry,the computing performance of super computer has been greatly improved.Sunway series is a heterogeneous supercomputer developed by China.However,because of its special architecture,there has been a lack of matching application software,so it is difficult to give full play to the computing performance of Sunway supercomputer.Numerical computing software occupies a large part in the application of scientific software.In the field of materials science,using first principles to calculate materials is the key to develop new materials.Through material calculation,the characteristics and behaviors of materials can be understood during material screening,which shortens the life cycle of material research and development and greatly reduces the cost of material research and development.In the calculation of materials,the traditional methods are density functional theory,molecular dynamics and computational fluid dynamics to calculate the physical and chemical properties of materials.Due to the rise of artificial intelligence,people use machine learning and deep learning models to model and analyze the material field,bringing new computational methods to the material field.Therefore,optimizing VASP on the new generation of Sunway supercomputer can not only play the computing power of Sunway supercomputer,improve the efficiency of material computing,but also help the cause of Chinese supercomputer,enrich the heterogeneous application library of Sunway supercomputer,improve the soft power of Chinese supercomputer.In addition,VASP software is studied and combined with emerging artificial intelligence to provide ideas for the subsequent research on material computing + artificial intelligence on Sunway supercomputer platform.In view of this,based on the new generation of domestic Sunway supercomputer,this thesis optimized the first-principles software VASP,including programming optimization and machine learning optimization.In addition,a crystal neural network algorithm model based on multi-head attention is proposed,and the predictive time of the model is compared with the traditional numerical calculation time.The main work and innovation points are as follows:(1)Based on the new generation of domestic Sunway supercomputer,first principles software VASP was transplanted and optimized for Sunway heterogeneous environment programming.Firstly,migration methods such as hybrid compilation and incremental compilation are used to enable it to call the slave core processor.Then,mathematical models are constructed to find the hot spot function of VASP and the objective function to be optimized in a quantitative way.Finally,optimization methods such as cooperative computation,double buffering and LDM level 2 cache are designed according to the software.In addition,an adaptive optimization method is proposed to solve the problem of negative optimization in the process of optimization,which can give full play to the computing performance of supercomputer.(2)Based on Bayesian inference method,a Bayesian incremental learning method suitable for VASP is proposed,which can optimize the calculation process of VASP.In order to solve the problem of inaccuracy caused by insufficient force field data when Bayes is constructing force field,incremental learning method is adopted to update and improve the prediction ability of the model by constantly adding new data and a small amount of model adjustment.At the same time,the model has the portability,which can be migrated to the environment where the model is inaccurate due to the small amount of calculation data.(3)Based on the crystal diagram neural network,a multi-head attention crystal diagram neural network(MH-CGCNN)model was proposed.This model used the multi-head attention mechanism to take the bond angles between different atoms in a molecule as the research object,and assigned weights corresponding to the actual macroscopic properties of molecules to different bond angles.Different weights reflected the degree of influence of the bond angles on molecular properties.In this way,we can simulate the interaction between the nearest neighbor atoms in real molecules.In addition,the bond Angle descriptor is introduced into the model,which makes the model have a better physical expression and reduces the mean absolute error MAE by 5.3%.