The Study Of Molecular Dynamics Parallel Simulation In Monocrystalline Silicon Grinding

Ultra-precision machining techniques of monocrystalline silicon have played an important role in the development of IC manufacturing. Ultra-precision grinding techniques are at the forefront of monocrystalline silicon ultra-precision machining technology. Conventional continuum-mechanics-based theory can not explain ultra-precision grinding mechanism and ultra-precision grinding experiments are difficult to be carried out and measured owing to size limitations. Molecular dynamics simulation (MDS) is able to simulate the ultra-precision grinding process on atomic scale and the processing parameters are easy to be acquired and analyzed, that makes it an important and effectual method to study ultra-precision grinding mechanism. However, the serial computing of MDS spends too much time to expand the simulation size. Parallel computing can decrease the simulation time effectively. So parallel computing's application to MDS will result in more accurate simulations and more intensive studies on ultra-precision grinding mechanism.First of all, a 3D MDS model of monocrystalline silicon grinding is established based on the analysis of the characteristics in ultra-precision grinding process. And MDS serial program has been investigated in detail in order to make parallel program come true. Then the paper design a MDS parallel algorithm in which domain is decomposed twice by use of PCAM (Partitioning, Communication, Agglomeration, and Mapping) combining the analysis of the grinding model. The parallel algorithms on different models are presented subsequently. By balancing between single transferring and package transferring of silicon atoms, package transferring and information visiting strategy based on "invariable sequence number" is designed. Then the concept of "atom relative list" is proposed to improve the listing method of atom neighbor list. The results show that "atom relative list" can reduce the computing time and simplify the program structure. Moreover, to ensure that the program running is not affected by the illegal break, the paper puts forward the approaches of saving processes and restarting program at the breakpoint. Besides, visual program for snapshots is programmed.In addition, through comparing results of MDS serial program and parallel program, the parallel program is demonstrated by the analysis of the snapshots of atoms' position and system energy. The accelerated ratio equal to 3.5 is achieved by running the parallel program on 4 processes and the efficiency reaches 87.5%. The simulation scale is expanded from kilo level of serial program to lakh level by optimizing model partitioning and increasingprocessor. Accordingly the computing time can be decreased by more than ten times. And the efficiency doesn't decline that proves the parallel algorithm is of good scalability.