Title: MiniBooNE Results worth waiting for
1MiniBooNEResults worth waiting for
- Heather Ray
- hray_at_fnal.gov
- Los Alamos National Laboratory
2Outline
- LSND MiniBooNE motivation
- MiniBooNE Experiment
- Why were waiting to open the box
- Improving the Optical Model
- Improving identification of mis-id ?0
- Particle ID Algorithm
3LSND The Great Mystery
- 1st accelerator expt to observe ? osc signal
- 3.8? excess of anti-?e in an anti-?? beam
- Incongruous with rest of osc results
- Other expt have explored LSND phase space but
allowed regions still remain
4MiniBooNE
Primary (protons)
Secondary (mesons)
Tertiary (neutrinos)
- 8 GeV proton beam
- 1.6 ?s pulse, 5 Hz rate from Booster
- p Be ? mesons
- Mesons focused by magnetic horn
- Mesons ? DIF ?
- E 500 MeV
- 800 Ton, 12 m diameter sphere
- Non-doped mineral oil
- Two regions
- Inner light-tight region, 1280 pmts (10
coverage) - Optically isolated outer veto-region, 240 pmts
5Why the Wait?
- The oscillation signal is expected to be small
- Probability for LSND oscillations 0.264!
- Need to know backgrounds, detector response very
precisely - Requires a well-developed, sensitive Particle ID
algorithm, exact optical model, solid
identification of mis-ID backgrounds
Why not borrow the optical model from another
mineral-oil based neutrino experiment?
6Why the Wait?
- No other ? expt uses non-doped mineral oil
- Were the first to study, model, and simulate ?
interactions in pure mineral oil - Scintillator fuzzes out rings, ruins separation
- SNO/Super-K H20, no fluor/scint, all Cerenkov
- LSND all scintillation (swamped fluorescence),
some Cerenkov - MB in the middle, need to untangle various
components
71st HurdleThe Optical Model
8The Optical Model
- Full battery of external measurements to provide
complete picture of OM - Problem! How do you set the relative
normalization from one measurement to the other?
(ie ratio of fluorescence to scintillation) - Need internal calibration sources / tank data to
provide correlations - We do not tune on any samples which may bias the
oscillation analysis
9External Measurements
- Variety of stand-alone tests which characterize
separate components of mineral oil
10Internal Calibration Sources
- Muon tracker cubes provides ? and Michel e-
of known position and direction in tank, key to
understanding E and reconstruction - Laser flasks (4) used to measure tube charge,
timing response - Neutral Current Elastic sample provides
neutrino sample, protons below Cerenkov threshold
isolate scintillation components, distinguish
from fluorescence of detector
11The Optical Model Chain
External Measurements and Laser Calibration
First Calibration with Michel Data
Calibration of Scintillation Light with NC Events
Final Calibration with Michel Data
Validation with Cosmic Muons, ?? CCQE, ?e NuMI,
etc.
12Recent Improvements
Improvements to OM greatly improve Michel
electron E as a function of
location in our detector
13Impact of Improved OM
Scintillation light in 1st gamma in pi0 fitter
Distance between pi0 vertex and 1st gamma
conversion point
142nd HurdleIdentifying Mis-IDs
15Minimizing Mis-IDs
- ?83 of all mis-ID backgrounds come from events
with a single ?0 - Need sample of pure ?0 to measure rate as
f(momentum) - High-P region very impt. to get a handle on
high-E ?e bgd from K
163rd HurdleParticle ID
17Sensitivity Estimate
- Good sensitivity requires PID
- Remove ? 99.9 of ?? CC interactions
- Remove ? 99 of all NC ?0 producing interactions
- Maintain ? 30-60 efficiency for ?e interactions
LSND best fit sin22? 0.003 ?m2 1.2 ev2
18Particle ID Algorithm
- Using a boosted decision tree
- Similar to a neural net, but better
- Needs to be trained on a set of variables
- Want vars which are powerful at distinguishing
between signal, background event types - Have a large list of potential inputs
- Require data MC shapes to agree for an input to
be considered for training - The more vars with agreement, the larger set of
powerful vars well have to draw from, thus
providing a more powerful PID algo
Nuc.Inst.Meth.A 543 (2005) 557-584 Nuc.Inst.Meth.A
555 (2005) 370-385
19PID Inputs
Calibration Sample
Signal-like Events
Primary Background
Mean 1.80, RMS 1.47 Mean 1.19, RMS 0.76
Mean 20.83, RMS 25.59 Mean 3.48, RMS 3.17
Mean 16.02, RMS 25.90 Mean 3.24, RMS 2.94
20Summary
- We are moving forward in leaps and bounds!
- Past 6 months have brought phenomenal improvement
in our Optical Model - Agreement in PID potential inputs vastly improved
- New pion fitter offers better resolution of
single ?0 events, reductions in mis-id
backgrounds - These improvements are vital to maximizing our
sensitivity to LSND - (Remember, Probability for oscillations 0.264)
- We are not done yet. Improvements are continuing
- hope to open box this summer
21BACKUP INFO
22NN vs Tree
The Elements of Statistical Learning, Hastie, Tibshirani, Friedman, Springer (2003) Neural Nets Trees
Natural handling of data of mixed type Bad Good
Handling of missing values Bad Good
Robustness to outliers in input space Bad Good
Insensitive to monotonic transformations of inputs Bad Good
Computational scalability (large N) Bad Good
Ability to deal with irrelevant inputs Bad Good
Ability to extract linear combinations of features Good Bad
Interpretability Bad Fair
Predictive power Good Bad
23Decision Trees
Pros
Cons
- Unstable - large trees have high variance
- Mitigate this by using a collection of trees
(boosting) - Dont capture additive structure well
- Use sensible choice of input vars
- Good Performance
- Low Bias
- Training is easy, does not depend on minimization
procedure - Immune to effects of outliers
- Resistant to effects of inclusion of irrelevant
input vars
24Why Boost a Tree?
- You can boost anything - tree, neural net, etc.
- Boosting combines weak classifiers to produce a
powerful committee - Classifiers are combined through a weighted
majority vote to produce the final output
25Boosted Trees
Cons
Pros
- Inherits pros of single trees
- Dramatic performance improvement
- Low bias, low variance
- Less susceptible to overtraining
- More of a black box
- Increases sensitivity to outliers and noisy data
26Boosted Tree Falsehoods
- Boosted trees are NOT robust against data to MC
disagreement - We must have good data to MC agreement for an
input to be used in training - Boosted tree performance does NOT improved with
the number of input variables
27Determining Backgrounds with MiniBooNE data
Full data sample 5.3 x 1020 POT
- High energy ne data
- Events below 1.5 GeV still in closed box (blind
analysis)
- ne from K
- Use High energy ne and nm to normalize
- Use Kaon production data for shape
- Need to subtract off misIDs
28Why the Wait?
- We dont have 2nd detector so we cant do flux
cancellation - We need to know the neutrino production
mechanisms much more precisely than past expts
have needed - Rely on data from external expts Harp thin
target results recently added to MiniBooNE MC
(April 06)
29Checking PID with NuMI Events
- Because of the off-axis angle, the beam at
MiniBooNE from NuMI is significantly enhanced in
nes from K - Enables a powerful check on the Particle ID
30Optical Model
- MB is very unique mineral oil with no
scintillator - Solar nu Genius Gd, Moon liq Ar, Heron
liq He, SNO heavy H20, Homestake Cl, Sage
Ga, Ge, Xe, GNO Ga, Gallex Ga, SuperK H20,
Borexino mineral oil PP0 (doped with a
fluor), ICARUS liq Ar - Reactor nu Chooz mineral oil Gd, Daya Bay
???, Diablo Canyon doped mineral oil, Kaska
???, Angra mineral oil Gd, Palo Verde ???,
Bugey ???, Gosgen ??? - SBL Accelerator expts Nomad collider detector
(drift chamber, etc), Chorus emulsifying film,
KARMEN liquid scintillator, LSND mineral oil
bPBD, NuTeV solid calorimeter, DoNUT
emulsion sheets - LBL Accelerator expts T2K ???, NoVa liquid
scintillator, MINOS solid detector, K2K H20,
Opera emulsion sheets
31Beams
- Nomad 450 GeV p Be
- Chorus 450 GeV p Be
- Karmen 800 MeV p heavy H20
- LSND 800 MeV p heavy H20
- NoVa 120 GeV p
- DoNUT 800 GeV p Tungsten