mcnp4c手册 1-36页：（我觉得下面的每一条都值得好好琢磨琢磨）
TIPS FOR CORRECT AND EFFICIENT PROBLEMS
This section has a brief checklist of helpful hints that apply to three phases of your calculation:
defining and setting up the problem, preparing for the long computer runs that you may require,
and making the runs that will give you results. Not everything mentioned in the checklist has been
covered in this chapter, but the list can serve as a springboard for further reading in preparation for
tackling more difficult problems.
A. Problem Setup
1. Model the geometry and source distribution accurately.
2. Use the best problem cutoffs.
3. Use zero (default) for the neutron energy cutoff (MODE N P).
4. Do not use too many variance reduction techniques.
5. Use the most conservative variance reduction techniques.
6. Do not use cells with many mean free paths.
7. Use simple cells.
8. Use the simplest surfaces.
9. Study warning messages.
10. Always plot the geometry.
11. Use the VOID card when checking geometry.
12. Use separate tallies for the fluctuation chart.
13. Generate the best output (consider PRINT card).
14. RECHECK the INP file (materials, densities, masses, sources, etc.).
15. GARBAGE into code = GARBAGE out of code.
B. Preproduction
1. Run some short jobs.
2. Examine the outputs carefully.
3. Study the summary tables.
4. Study the statistical checks on tally quality and the sources of variance.
5. Compare the figures of merit and variance of the variance.
6. Consider the collisions per source particle.
7. Examine the track populations by cell.
8. Scan the mean free path column.
9. Check detector diagnostic tables.
10. Understand large detector contributions.
11. Strive to eliminate unimportant tracks.
12. Check MODE N P photon production.
13. Do a back-of-the-envelope check of the results.
14. DO NOT USE MCNP AS A BLACK BOX.
C. Production
1. Save RUNTPE for expanded output printing, continue run, tally plotting.
2. Look at figure of merit stability.
3. Make sure answers seem reasonable.
4. Make continue runs if necessary.
5. See if stable errors decrease by 1/sqrt(N) (that is, be careful of the brute force approach).
6. Remember, accuracy is only as good as the nuclear data, modeling, MCNP sampling
approximations, etc.