In the previous section it was demonstrated how following just a few puffs, representing the growth of the particle distribution with time, gave comparable results to following the individual particles. However, one limitation of the puff approach is that the meteorological conditions used for the transport and dispersion are valid only at the puff center position. When the puff expands beyond a few meteorological grid points, one data point can no longer be used to represent the puff. In this situation, a single puff is split into multiple smaller puffs, each with a proportional fraction of the original mass.
 The puff splitting process occurs in the horizontal when the puff grows larger that the meteorological grid size. To more clearly see this effect, go back to the 1 particle release calculation of the previous example
(gauss_control.txt and gauss_setup.txt) and rerun the calculation. The simulation at 12 hours shows the effect of the horizontal splitting, particularly evident by the circular puffs near the edge of the main pollutant cloud. Slight differences may exist between your calculation and the illustration.
 The exact splitting times can be determined from the diagnostic message file. Press the Advanced / View MESSAGES menu tab to open the listing MESSAGE file. The lines starting with NOTICE main: give the simulation hour, the time in minutes, the number of particles/puffs, and the total mass of all the particles on the computational domain. The first split occurs after hour 7, when the calculation goes from 1 to 5 puffs. Note the total mass stays the same, at 1.0 unit, now distributed over 5 puffs. At the end of the calculation, we are following 129 puffs.
 In a Gaussian simulation, one puff splits into five when the puff size exceeds the meteorological grid size. Therefore, the number of puffs increases at a rate of 5^{(number of splits)} so that the sequence would be: 1, 5, 25, 125, ..., but as shown in the MESSAGE file the number is actually: 1, 5, 9, 33, 57, 121. ... because not all puffs grow at the same rate, puff merging occurs when previously split puffs are in the same location, and some puffs may exit the computational domain (WRF grid). So there is no simple formula to compute the puff number growth rate.
The use of puff splitting improves the representativeness of individual puffs at longer transport distances. However, puff splitting also results in an ever increasing number of puffs in the computation, resulting in longer simulation times. When the number of puffs approaches the maximum allocated array space (MAXPAR), first puff splitting is turned off. A more serious concern is when this occurs during a continuous emission simulation because emissions may be turned off if no array space is available. This event is always noted in the MESSAGE file but may easily be overlooked. This could result in a substantial concentration underprediction and can only be addressed by increasing the value of MAPPAR.
