Difference between revisions of "SoLID Ecal Weekly 20230803"
From solidwiki
(Created page with "==Update from Carter on Timing== * Reminder: SC_C - SC_D yields about 450ps. Rough estimate based on scintillator thickness gives about 210ps for C-D. So we are ~ factor 2 off...") |
(→Update from Darren on ML PID) |
||
| Line 6: | Line 6: | ||
==Update from Darren on ML PID== | ==Update from Darren on ML PID== | ||
| − | * | + | *[https://solid.jlab.org/wiki/images/c/c0/Week9_Tuesday.pdf Darren's Slides], summary: |
**with higher statistics for background, completed beamtest sim + bg (x1, x10) ML study. Typical charged pion rej is 1E3 and typical pi0 rejection is 8-10, both at an electron efficiency of 95%. The "pi0 vs. e" (corresponding to 2x2 matrix, matrix not shown) and the "pi+/-, pi0, e" (3x3) show similar results for pi0, though the 3x3 matrix is slightly worse. | **with higher statistics for background, completed beamtest sim + bg (x1, x10) ML study. Typical charged pion rej is 1E3 and typical pi0 rejection is 8-10, both at an electron efficiency of 95%. The "pi0 vs. e" (corresponding to 2x2 matrix, matrix not shown) and the "pi+/-, pi0, e" (3x3) show similar results for pi0, though the 3x3 matrix is slightly worse. | ||
**studied TS1, | **studied TS1, | ||
Revision as of 15:19, 1 August 2023
Update from Carter on Timing
- Reminder: SC_C - SC_D yields about 450ps. Rough estimate based on scintillator thickness gives about 210ps for C-D. So we are ~ factor 2 off (assuming similar PMT TTS as HRS S2 and S2m)
- (T+B)/2 for LASPD minus SC_C (or minus SC_D) gives about 750ps in sigma, or under 700ps with tight tracking cuts in x. If using 150ps (LASPD) plus 62.5ps (edge finding) plus 450ps/sqrt(2) (for either C or D alone), we expect 340ps. So we are factor 2 off.
- discussion of drift distance: C (56.2mm) LASPD (68.9mm) D so total about 150mm from C to D counting LASPD thickness (2cm).
- Suggested to do: try (T+B)/2 minus (C+D)/2 to see what the sigma value is, then plot this (or -C or -D as before) vs. tracking y to see if there is any correlation to be corrected.
Update from Darren on ML PID
- Darren's Slides, summary:
- with higher statistics for background, completed beamtest sim + bg (x1, x10) ML study. Typical charged pion rej is 1E3 and typical pi0 rejection is 8-10, both at an electron efficiency of 95%. The "pi0 vs. e" (corresponding to 2x2 matrix, matrix not shown) and the "pi+/-, pi0, e" (3x3) show similar results for pi0, though the 3x3 matrix is slightly worse.
- studied TS1,
- To do: Given the limited time left, will focus on getting the new pi0 simulation read correctly
- Moving forward regarding applying ML to actual data -- Darren will give the code to Zhiwen. Two possibilities:
- use the "clean high-energy e sample" and "clean pi0 sample" (from Ye) and "clean pion sample" (from Mike) to both train and test the ML, but this may not be worth doing because we expect "high performance".
- use the ML classifier trained by beam test sim + bg on the two clean samples and see how they work.
- use the ML classifier trained by beam test sim + bg on the full data set and see how they work.
Update from Ye on sim
From simulation meeting:
- eAll code used phi (should be cos(phi)) for event generation. This is missing sin(phi) factor and resulting in higher yield than the correct value. The difference is small at forward angles -- sin(phi)/phi, but large for phi close to 90deg. For PVDIS configuration, 22-35 deg, the rates are 30% higher (than real value). This eAll sim was used in all 2022 PVDIS impact studies, including PDF dependence and 11 vs. 22 GeV comparison, as well as BNSSA proposal. However, note that the study also missed radiative factor which would cause rates to be 10-20% higher (due to resonances having higher rates). So effects on the rate estimation is at 10-20%.
