基于并行化智能优化算法的材料大数据处理研究

发布时间：2018-11-25 17:07

【摘要】：大数据时代的来临改变了人们生活的方方面面,也为材料科学的发展带来了新的机遇和挑战。随着材料基因组计划的提出,材料大数据处理研究已经成为一个新的社会热点。本文就针对材料大数据处理中的分子力场优化问题进行了深入研究。智能算法又称为元启发式算法,在解决此类问题中有突出表现,其中遗传算法和粒子群算法应用尤为广泛。随着社会和科学的发展,各个领域中的数据都越来越多,也越来越复杂,本文要解决的主要问题,分子力场优化问题就是一个例子,另外,很多实际应用还对实时性有要求。因此,串行智能算法已经难以满足应用需求,对之进行并行化研究十分必要。本文在深入分析智能算法本质特征的基础上,实现了智能算法主要是遗传算法和粒子群算法的并行化处理。其流程大致为,第一,利用福斯特并行算法设计方法对遗传算法和粒子群算法进行了特征分析和并行化设计,制定了主从式遗传算法并行化设计策略和并发主从混合式粒子群算法并行化设计策略。第二,利用当下流行的OpenMP应用程序接口对遗传算法和粒子群算法的计算密集部分,即适应度评价函数进行了并行化处理,实现了智能算法的主从式并行化,使得算法的运算速度随着所用核数的增加成比例提升。第三,利用MPI消息传递接口混合OpenMP应用程序接口对划分子群的粒子群算法进行了并行化处理,进一步提升算法执行速度的同时,也使算法的优化效果得到了很大的提升。其中MPI用于子群信息交互,OpenMP仍然用于适应度函数并行化处理。另外,针对基于子群划分的并行化粒子群算法,本文提出一种交流子群个体历史信息和加入遗传机制的策略,提升了算法的收敛速度,实验表明,对于时间复杂度很高的问题,该算法可以达到更好的优化效果。最后,以核用辐照碳化硅为例,本文将并行化智能算法应用到了材料大数据中的分子力场优化问题中。实验证明,并行化智能算法有更低的时间复杂度和更高的优化性能,而且针对此类时间复杂度很高的实际问题,本文提出的加入子群个体历史信息和遗传机制的基于子群划分的并行化粒子群算法有很好的表现。
[Abstract]:The coming of big data has changed all aspects of people's life and brought new opportunities and challenges to the development of material science. With the development of material genome project, the study of material big data has become a new social hotspot. In this paper, the optimization of molecular force field in the treatment of material big data is studied. Intelligent algorithm, also known as meta-heuristic algorithm, has outstanding performance in solving such problems, especially genetic algorithm and particle swarm optimization algorithm. With the development of society and science, the data in various fields are more and more complex. The main problem to be solved in this paper is the optimization of molecular force field. In addition, many practical applications require real-time performance. Therefore, the serial intelligent algorithm has been difficult to meet the application requirements, and it is necessary to research on parallelization. On the basis of analyzing the essential characteristics of intelligent algorithm, this paper realizes the parallel processing of genetic algorithm and particle swarm optimization algorithm. The flow chart is as follows: first, the genetic algorithm and particle swarm optimization algorithm are analyzed and parallelized by using the Foster parallel algorithm design method. The parallel design strategy of master-slave genetic algorithm and concurrent master-slave hybrid particle swarm optimization algorithm is proposed. Secondly, the popular OpenMP application program interface is used to parallelize the computation dense part of genetic algorithm and particle swarm optimization, that is, fitness evaluation function, and realize the master-slave parallelization of intelligent algorithm. The computation speed of the algorithm increases proportionally with the increase of the number of kernels used. Thirdly, the hybrid OpenMP application program interface of MPI messaging interface is used to parallelize the particle swarm optimization algorithm, which further improves the execution speed of the algorithm, and also improves the optimization effect of the algorithm greatly. MPI is used for subgroup information interaction and OpenMP is still used for parallelization of fitness function. In addition, for the parallel particle swarm algorithm based on subgroup partition, this paper proposes a strategy of exchanging historical information of subgroup and adding genetic mechanism, which improves the convergence speed of the algorithm. Experiments show that, for the problem of high time complexity, The algorithm can achieve better optimization effect. Finally, taking nuclear irradiated silicon carbide as an example, the parallel intelligent algorithm is applied to the optimization of molecular force field in big data. Experiments show that the parallel intelligent algorithm has lower time complexity and higher optimization performance. The parallel particle swarm optimization algorithm based on subgroup partition proposed in this paper has a good performance by adding historical information and genetic mechanism of subgroups.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP18;TP311.13

【参考文献】