Accelerating DSMC data extraction.

Gallis, Michail A. ; Piekos, Edward Stanley

Sandia National Laboratories

关键词: 99 General And Miscellaneous//Mathematics, Computing, And Information Science;    Gas Flow-Mathematical Models.;    Monte Carlo Method;    Computer Calculations;    Optimization Stochastic Processes.;   

DOI  :  10.2172/922066
RP-ID  :  SAND2006-6692
RP-ID  :  AC04-94AL85000
RP-ID  :  922066

美国|英语

来源: UNT Digital Library

【 摘 要 】

In many direct simulation Monte Carlo (DSMC) simulations, the majority of computation time is consumed after the flowfield reaches a steady state. This situation occurs when the desired output quantities are small compared to the background fluctuations. For example, gas flows in many microelectromechanical systems (MEMS) have mean speeds more than two orders of magnitude smaller than the thermal speeds of the molecules themselves. The current solution to this problem is to collect sufficient samples to achieve the desired resolution. This can be an arduous process because the error is inversely proportional to the square root of the number of samples so we must, for example, quadruple the samples to cut the error in half. This work is intended to improve this situation by employing more advanced techniques, from fields other than solely statistics, for determining the output quantities. Our strategy centers on exploiting information neglected by current techniques, which collect moments in each cell without regard to one another, values in neighboring cells, nor their evolution in time. Unlike many previous acceleration techniques that modify the method itself, the techniques examined in this work strictly post-process so they may be applied to any DSMC code without affecting its fidelity or generality. Many potential methods are drawn from successful applications in a diverse range of areas, from ultrasound imaging to financial market analysis. The most promising methods exploit relationships between variables in space, which always exist in DSMC due to the absence of shocks. Disparate techniques were shown to produce similar error reductions, suggesting that the results shown in this report may be typical of what is possible using these methods. Sample count reduction factors of approximately three to five were found to be typical, although factors exceeding ten were shown on some variables under some techniques.