Vetting the Low EM physics?

    It is becoming clear that the Low EM option has nontrivial effects on the simulated physics of STAR.  A previous study showed the difference between sampling fractions in nominal GEANT and that tuned with the Low EM option.  Before adopting Low EM as the default, however, the physics must be validated.  Is GEANT getting the low energy physics correct?  Which sampling fraction truly describes the detector?

    Any conclusive answer would have to come from detailed data/simulation comparisons for variables sensitive to the low energy physics.  Aside from modifying the sampling fraction, the Low EM option does not seem to have an effect on the towers themselves.  This really just leaves the SMD response.

    Unfortunately, comparable data and simulation are not immediately available.  Data and simulation from 2006 exist, but the 2006 geometry used in the simulations is known to be incomplete.  Moreover, neither is ready for immediate analysis due to significant upgrades to the analysis code in the time since the relevant trees were produced.  Some data has been produced for 2009, but full calibration of the SMD is not yet ready and sufficiently large simulation samples are far from available.  To motivate the use of Low EM, then, we're going to have to risk the health of our wrists do some hand waving.

    One of the known discrepancies between data and simulation in 2006 is the shower shape in the SMD.  The showers themselves are narrower in simulation than in data.  Likewise, the total number of strips above pedestal is smaller in simulation than in data.  Now if the Low EM option could improve these characteristics, widening showers and increasing the number of active strips, we would have some motivation to believe in Low EM.

    Below are the results from two particle simulation studies.  Each sample consisted of 105 single photons thrown in a small fiducial volume around the center of a single tower, one with the nominal GEANT tune and the other with Low EM enabled.  Low EM does indeed have the desired effect, increasing the RMS (and hence width) of each shower along with the number of active strips.

 

 

    So where does that leave us?  The anecdotal evidence suggests that Low EM is a step in the right direction.  The new tune is clearly not doing anything crazy, so the argument just has to be made to justify the increased computation time.

    Let's play with some numbers.  Assuming that the nominal tune requires 8 s to process one event (1s for GEANT, 7 s for BFC), switching to Low EM would require a 40% increase in computation time.  Weeks stretch to months, but there is no difference in overall footprint.  Taking a conservative approach, let's say that all of the results so far imply a 25% probability that Low EM will be necessary for the 2009 simulation productions.  Then the expected computation time for the nominal sample is

p(nominal good enough) * t(nominal) + p(low EM needed) * [ t(low EM) + t(nominal) + t(diagnose) ]

p(nominal good enough) * t(nominal) + p(low EM needed) * [ 1.4 * t(nominal) + t(nominal) + t(diagnose) ]

0.75 * t(nominal) + 0.25 * [ 2.4 * t(nominal) + t(diagnose) ]

1.35 * t(nominal) + 0.25 * t(diagnose)

whereas the expected time for running with Low EM to begin with is

1.4 * t(nominal)

If it will take anywhere near as much time to diagnose a need for Low EM (waiting for data production to be finished, tuning calibrations, updating trigger emulation, running trees, etc) as it does to run the simulation itself then we might as well just take the hit now.  Especially if we could sneak the simulation production in before data begins to be produced.