Various Rupak Mahapatra (for Angela, Joel, Mike & Jeff) Timing Cuts

Different Timing Cuts 2-D pminrtc & pdelc: Angela 2-D pfracc & pdelc: Rupak/Joel 4-D  2 pfracc, pminrtc, pdelc & y: Rupak/Joel 2-D Neural Net pfracc & pdelc: Mike A Technique presentation (  2 ): Jeff

General Idea Nuclear Recoil Band based on Yield Recoil Energy (keV) Z2/Z3/Z5 20 keV 10-40% Z Reduced ionization collection and fast rise time

The Discriminators: Use Combination Yield = Ionization Energy/ Phonon Recoil Energy Surface recoils tend to have lower ionization, mix with NR Pfracc = Highest_Phonon_Energy/Opposite_Phonon_Energy Surface recoils on phonon side tend to have higher energy partition due to proximity to phonon sensors pminrtc = 10%-40% risetime of the largest phonon signal Surface recoils tend to have lower risetime than bulk recoils pdelc = 20% delay of largest phonon signal wrt to charge st. Surface recoils tend to have lower delay than bulk recoils All discriminators are correlated to some degree All discriminators have energy dependence to some degree

Definition of Surface Recoils(  ) Low Yield events from Ba R118: reject  in 3  NR R119: Use wider definition  : 0.1 < yic <  - 5  More stats for cut defn. Establish timing cut to reject all but “n” wide  s Estimate closed leakage based on allowed leakages in open dataset

2-D cut in pminrtc and pdelc: Angela Define delpric and rtpric: energy corrected pdelc and pminrtc For events between 10 and 100keV in pric, plot pdelc and pminrtc neutrons and betas. Fit delpric-rtpric distributions for neutrons with gaussian, exclude all events (betas) outside 4 sigma of this distribution. Define a cut in delpric+rtpric that allows desired beta leakage (set on the sixth event in each detector for this analysis.

Issues associated with timing cuts  T2Z5 charge collection: Walter did a study of the timing outliers in T2Z5, which are all clustered at the bottom of the delay plot. He defined several cuts to exclude the affected region. In this analysis, the cut used is a simple ydel>-20 cut.  T1Z1has low efficiency as usual

Results Allowed 5 events of leakage in each detector for an exected leakage of 6 events/detector*6Ge detectors =36 betas. In the WIMP search data, the leakage from these cuts is expected to be on the order of a fraction of an event overall. Neutron efficiencies for these cuts are around 75% in the higher energy bins and worse at low energies. Leakage by E bin

Low energies and timing  Below 7keV, Long found uncertainty in charge energy, causing leakage from ER toNR  Below 15keV, uncertainty in Qist that affects pdelc. Mostly does not affect the timing cut  Below 10keV, ER and NR bands not well separated, hence difficult to define  population  Below 20keV, timing cut efficiency is low ~ around 20%  Further work needed to determine where to set our analysis threshold. Probably it should be above 7keV.

2-D pfracc & pdelc: Rupak/Joel Reject all  in 3  NR Found 9 extreme outliers. 3 rejected in T2Z5 by Walter’s cut Discovered that flash time cut may be helpful Flashtime RQ broken. Joel defined a function that utilizes DAQ gpib log to get flashtime Efficiency ~ 70-80% w 6 leakages in open Ba

4-D y,pfracc,pminrtc, pdelc: Rupak/Joel Similar to Vuk’s rb-rn method, normalized to 1 Define 4-D space in y, pfracc, pminrtc & pdelc Calculate neutron and  centroids from 3  NR In 4-D space, calculate dist. of each event from neutron (rn) and  (rb) Define cut to reject all  Efficiency ~ 70-80% w 1 leakage in open Ba

4-D Cut with  2 : Rupak/Joel Earlier 4-D cut assumes each discriminator has equal power in discr. Assign natural weights to each discriminators based on their accuracy in discrimination Calculate combined  2 distance for each event from neutron and  Efficiency > 80% with 1 leakage in open Ba with WIDE  distribution

R119 pfracc & pdelc Timing Cuts using a Neural Network M. Attisha Can create cuts that would be v. tough to parametize by hand Easy to experiment with a range of input (RQ) parameters Cut must be chosen based on performance upon simulation data Training Data = Ba open + Cf (even Event#) Simulation Data = Ba closed + Cf (odd Event#) Trained on all betas & neutrons within the 3σ NR band The cut is calculated in each ZIP for a single energy bin (pric): keV

~90% rejection below 60keV pric Reduced efficiency at higher energies due to low neutron stats 6 beta leakage events remain in Z2, Z3, Z5, Z9 & Z11 Current rejection performed using pdelc and pfracc Naïve addition of other parameters such as pminrtc gives little improvement, but plan to study effect of weighting inputs R119 Timing Cuts using a Neural Network, cont. M. Attisha

Chi-squared Methods: Jeff Our pulse shape and timing parameters show differing discrimination powers and significant correlations –We’d like to figure out the appropriate “metric” on this multi-dimensonal parameter space. In the case of gaussian parameters, the optimal metric is provided by the inverse of the covariance matrix Basic Plan: –Preselect samples of neutrons and betas (and perhaps gammas), calculate mean, cov for each population –Compute –Make a 1-D cut in –Reject outliers with large chi-squareds

Chi-Squared Methods Do we include yield? –Most rejection power, but betas clearly non- gaussian –For now, separate band cut Energy dependence –Work in energy bins at first –Use energy variations in computed mu, sigma to define energy corrections –Perform energy- independent cut (or fewer bins, at least)

Status and Plans Standard 2-D timing cut looks polished 4-D timing cut needs refinement to take full advantage of correlations for better rejection Not clear whether to include the best discriminator y in likelihood. Excluding gives easier leakage estimate Neural Net analysis looks promising Important to fully exploit all discrimination parameters provided by our tremendous detectors Timing cut group needs to formulate an action plan