Most DSMC programs continue to employ outdated procedures.
The most extensive description of the DSMC procedures is that in Molecular Gas Dynamics and the Direct Simulation of Gas Flows (G. A. Bird) Oxford University Press, 1994. Most of these procedures have been largely superseded over the past fifteen years. The newer developments include separate collision and sampling cells, optional nearest-neighbor collisions, variable adaptive time steps as well as cell sizes, modification of the NTC collision routine to avoid the use of averaged quantities and to remove the ambiguity if there is only one simulated molecule in a collision cell, and the automatic setting of all computational parameters.
It has been shown [M. A. Gallis, J. R. Torczynski, D. J. Rader and G. A. Bird, J. Computat. Phys. v. 228, p. 4532, (2009)] that the rate at which a result converges to the correct value with increasing molecule number and sample size is much greater for programs that employ these “sophisticated” procedures than it is for programs that employ the older “simple” procedures. An fuller description of the new procedures has been provided in the notes prepared for the DSMC07 meeting in Santa Fe.
The sophisticated procedures have been implemented in the DS1V and DS2V programs that appear in dedicated pages on this site. Note that source code is provided for the DS1V program. The procedure that has the most beneficial effect in terms of the number of simulated molecules that are required for a converged calculation is the use of nearest-neighbor collisions. Its use is optional because it has been associated with two computational artifacts that do not appear when it is disabled and are completely uncharacteristic of DSMC calculations. These are a spike in the heat transfer that sometimes occurs at the axis in hypersonic blunt-body calculations in DS2V, and blips that sometimes occur in the flow upstream of a shock wave in DS1V. Dr. Michael Gallis (Sandia National Labs.) has shown that the latter effect is associated with excessively large time steps and it is hoped that automatic fixes will be developed. For the moment, care should be taken when the nearest-neighbor collision option is selected.
I often re-calculate cases that appear in current papers using the DS2V program. I continually find that the results of calculations that have have employed many millions of molecules, generally on parallel clusters, can be reproduced with about 30,000 molecules. The DS2V calculations are made in a few hours on an ordinary laptop PC. There are no differences in the results and the internal checks within DS2V report that the criteria for a good calculation have been met.
Most DSMC programs are difficult to use and/or are not generally available.
Many contemporary papers continue to employ custom programs that are dedicated to just one class of problem. This type of program can be written in about one month if it employs simple procedures or about three months if it employs sophisticated procedures. A general program requires about one man-year if it has a minimal level of integrated data input and graphical output and two years for a program with a “commercial grade” interface. These are minimum times and DS2V was developed over ten years of part-time effort that probably amounted to two or three man-years. This program should now be replaced by a fully 64 bit parallel program with similar versatility and usability. At the same time, it remains the most user-friendly program and has been successfully applied by many workers with little or no background in either DSMC or gas dynamics. It is arguably the most efficient program and, being readily available, it provides a benchmark against which the existing and new programs can be assessed. The foamDSMC program is an interesting new initiative, but definitive benchmark results are not yet available.
The optimum geometry model and parallelization schemes are not yet clear.
The basic geometry choice is between schemes that employ a rectangular reference scheme and those that employ ray-tracing schemes. With rectangular schemes, there is a choice between multi-level homogeneous grids (as in DS2V/3V) and “tree-structure” grids. Homogeneous grids are computationally fast, scale easily from two to three dimensions and allow extremely rapid adaption to the optimum number of molecules in both collision and sampling cells. Ray-tracing schemes have the doubtful advantage of using well-established continuum CFD grids and are relatively slow.