1990 Fiscal Year Final Research Report Summary
Expert System of Automatic Programming for the Monte Carlo Direct Simulation
| Project/Area Number |
63550134
|
|---|
| Research Category |
Grant-in-Aid for General Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
Fluid engineering
|
|---|
| Research Institution | Mie University |
|---|
Principal Investigator |
USAMI Masaru Mie University, Faculty of Engineering, Associate Professor, 工学部, 助教授 (10106974)
|
|---|
| Project Period (FY) |
1988 – 1990
|
| Keywords | Rarefied Gas / Direct Simulation / Monte Carlo Simulation / Vector Processing / Parallel Processing / Neural Network / Automatic Programming |
|---|
| Research Abstract |
The direct simulation Monte Carlo (DSMC) method costs a lot of CPU time, especially for the near-continuum regime. To use supercomputers with vector processors in the most effective way, a computer program must be adequately vectorized. By frequent use of a programming technique called "data collection" a simulation program eight times faster than a conventional one is obtained. The DSMC always deals with an unsteady flow, and the time appearing in the simulation has close connection with real time. In heat transfer through a rarefied gas from a solid body, the heat capacity of the body is prohibitively larger than the quality of heat transferred by gas molecules. In that situation two types of characteristic time are required. To connect these different characteristic times, a weighting factor for the time scale is introduced. The efficiency of the parallel processing is investigated. The INMOS T800 transputer has four communication links that allow networks of transputers with high communication speed. Each transputer covers its own region of the simulated physical space and the information of molecules are transferred through the high speed link when they cross the boundary of the regions. The double structure of cells is a very flexible, powerful scheme for rapidly cell identification. In that, the flow field is first divided by very small rectangular grid-cells (sub-cells). Since the macro-cell in which intermolcular collisions are counted and the macroscopic properties are sampling is constructed of many grid-cells, the simulated region of physical space can be divided freely into a network of macro-cells. A neural net is investigated to obtain a table indicating which macro-cell each grid-cell belongs to. Automatic programming is now in progress for the DSMC method.
|
