基于GPU的颗粒增强复合材料损伤演化的宏细观跨尺度模拟

发布时间：2018-05-14 19:46

本文选题：颗粒增强复合材料 + CUDA　；参考：《昆明理工大学》2015年硕士论文

【摘要】：颗粒增强复合材料作为一种力学性能优越的新型材料,已经被广泛地应用于航空航天、机械制造和光学,,逐步地替代了一些传统的金属材料。由于在基体中加入增强颗粒,颗粒增强复合材料的多项力学性能均优于基体材料本身,与此同时,增强项的加入又降低了材料的疲劳断裂性能。近年来,随着颗粒增强复合材料越来越多地出现在日常生活的各个领域,国内外大量的研究者针对该材料进行了一系列的数值模拟工作。随着科学和技术的发展,针对一些复杂的模型的精确模拟,研究人员需要解决计算耗时长和占用内存量大这两大主要问题。本文根据GPU的可编程能力和高速并行处理能力的优势,以实现并行计算加速为出发点,将CUDA编程融入到颗粒增强复合材料损伤演化的宏细观跨尺度模拟。实现了CUDA Fortran程序和Fortran程序的混合编程。简要分析了并行程序与串行程序的区别的同时,突出了GPU加速的合理性、可行性和优越性。并且着重分析了CUDA编程要点以及充分利用共享存储器Share Memory的重要性,不仅能够很好的隐藏了数据传输延迟,而且也在一定程度上节省了存储空间和时间,从而提高了效率。利用CUDA编程技术,改进了原来计算程序的算法流程。在原程序计算得到宏观有限元位移解,整个构建离散为细观模型尺度与宏观大范围尺度之后,针对Voronoi单元的单元刚度矩阵计算时调用CUDA程序进行计算,重新编写了相应的Voronoi单元的单元刚度矩阵的程序。最后将最终的结果与宏观元计算结果耦合,得到整个模型的应力应变响应。建立多个计算模型,模拟颗粒增强复合材料损伤演化的过程。并分析模型在不同实验环境、不同高斯点数量、不同应力参数情况下的计算速度,并通过与原程序计算结果对比进行充分的验证。其主要结论为计算的数据规模较小时,GPU并行计算得到的加速并不明显；当数据规模越大时,GPU的加速效果越明显；CUDA核心越多,得到的提速效果越好。
[Abstract]:As a new material with superior mechanical properties, particle reinforced composites have been widely used in aerospace, mechanical manufacturing and optical materials, gradually replacing some traditional metal materials. Because of the addition of reinforced particles in the matrix, the mechanical properties of the particle reinforced composites are better than that of the matrix material itself, and at the same time, the fatigue fracture properties of the composites are reduced by the addition of the reinforced particles. In recent years, with the increasing emergence of particle reinforced composites in various fields of daily life, a large number of researchers at home and abroad have carried out a series of numerical simulation work on the material. With the development of science and technology, researchers need to solve the two major problems of long computation time and large amount of memory for the precise simulation of some complex models. According to the advantages of GPU's programmable ability and high speed parallel processing ability, this paper integrates CUDA programming into macro and mesoscale simulation of damage evolution of particle reinforced composite materials, taking the acceleration of parallel computing as the starting point. The mixed programming of CUDA Fortran program and Fortran program is realized. The difference between parallel program and serial program is briefly analyzed, and the rationality, feasibility and superiority of GPU acceleration are highlighted. The importance of CUDA programming and making full use of shared memory Share Memory is analyzed. It not only hides the data transfer delay, but also saves storage space and time to a certain extent, thus improving the efficiency. The algorithm flow of the original program is improved by using CUDA programming technology. After the macro finite element displacement solution was obtained by the original program, and the whole construction was discretized into mesoscopic model scale and macro large scale, the CUDA program was used to calculate the element stiffness matrix of Voronoi element. The program of element stiffness matrix of Voronoi element is rewritten. Finally, the stress and strain response of the whole model is obtained by coupling the final result with the result of macro element calculation. Several computational models were established to simulate the damage evolution of particle reinforced composites. The calculation speed of the model under different experimental conditions, different Gao Si points and different stress parameters is analyzed, and the results are compared with those of the original program. The main conclusion is that the acceleration of GPU parallel computation is not obvious when the data scale is small, and when the data scale is larger, the acceleration effect of GPU is more obvious than that of CUDA core, and the speedup effect is better.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TB33

【参考文献】