Machine Learning workflows in KM3NeT
Transcript of Machine Learning workflows in KM3NeT
![Page 1: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/1.jpg)
Machine Learning workflows in KM3NeT
Stefan Reck
IWAPP workshop
2021-03-10
![Page 2: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/2.jpg)
ORCA and ARCAsame technology – different sites
![Page 3: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/3.jpg)
3Machine Learning workflows in KM3NeT
KM3NeT: neutrino detector network
ORCA
![Page 4: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/4.jpg)
4Machine Learning workflows in KM3NeT
KM3NeT: neutrino detector network
ARCA
![Page 5: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/5.jpg)
5Machine Learning workflows in KM3NeT
→ detection principle: measure Cherenkov radiation of charged particles
43 cm
KM3NeT: neutrino detector network
DOM
DOM 31 PMTs
String 18 DOMs
Total (planned) 3 blocks of 115
strings each
Currently 6 strings (ORCA)
1 string (ARCA)
![Page 6: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/6.jpg)
6Machine Learning workflows in KM3NeT
KM3NeT ORCA
KM3NeT ARCA _
- for neutrinos with TeV energy
- cosmic neutrinos, …
- height: 600m
- line spacing: about 90m
- for neutrinos with GeV energy
- neutrino oscillations, mass ordering, …
- height: 150m
- line spacing: about 20 - 23m
KM3NeT: neutrino detector networkhttps://arxiv.org/abs/1601.07459
![Page 7: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/7.jpg)
7Machine Learning workflows in KM3NeT
Neutrino interactions
νμ − CCνe − CC ντ − CC ν − NC
Credit: J. Tiffenberg,
NUSKY11
![Page 8: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/8.jpg)
8Machine Learning workflows in KM3NeT
Detector signatures
shower-like shower-like (83%) shower-like
track-like (17%)
track-like
Credit: J. Tiffenberg,
NUSKY11
![Page 9: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/9.jpg)
9Machine Learning workflows in KM3NeT
● Two types of background producing photons in the deep sea:
1. Atmospheric muons passing the detector from above
2. Random noise, by K-40 beta decays and bioluminescent organisms
μ μ μ
νμ − CC
KM3NeT backgrounds
![Page 10: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/10.jpg)
10Machine Learning workflows in KM3NeT
𝝁
𝝂𝝁
Up-going ν𝜇 − CC track-like event ν𝑒 − CC shower-like event
𝝂𝒆
Event topologieshttps://arxiv.org/abs/1601.07459
![Page 11: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/11.jpg)
11Machine Learning workflows in KM3NeT
𝝁
𝝂𝝁
Up-going ν𝜇 − CC track-like event ν𝑒 − CC shower-like event
𝝂𝒆
Event topologies
How to separate these topologies?
![Page 12: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/12.jpg)
Random Decision
Forests
![Page 13: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/13.jpg)
13Machine Learning workflows in KM3NeT
Random decision forest
▪ ensemble of decision trees
▪ input: "hand-crafted" features
→ these need to be manually designed
▪ each tree is trained on random subset of features
▪ each tree outputs either "track" or "shower"
→ final score:
in total: ~150 features!
![Page 14: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/14.jpg)
14Machine Learning workflows in KM3NeT
what are the input features?
→ e.g. fit quality
▪ use maximum-likelihood-based reconstruction of observables
▪ compare the reco quality of track and shower hypothesis
Random decision forest
![Page 15: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/15.jpg)
15Machine Learning workflows in KM3NeT
what are the input features?
→ e.g. hit distribution
▪ compare distribution of hits to expectation in simulations
▪ shower events are more spherical
Random decision forest
![Page 16: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/16.jpg)
16Machine Learning workflows in KM3NeT
Random decision forest
Result:
▪ good separation between track-
like and shower-like events
▪ define separation power S:
quantifies the overlap
between the distributions
PhD thesis Steffen Hallmann
![Page 17: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/17.jpg)
17Machine Learning workflows in KM3NeT
Random decision forest
Result:
▪ good separation between track-
like and shower-like events
▪ define separation power S:
quantifies the overlap
between the distributions
RDFs are also used for separating
neutrinos from background
PhD thesis Steffen Hallmann
/ORCA
![Page 18: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/18.jpg)
18Machine Learning workflows in KM3NeT
hits
features
track/shower?
extract
RDF
difficult
incomplete?
● Problem: feature design is not easy and maybe we missed
some good features?
Random decision forest
![Page 19: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/19.jpg)
19Machine Learning workflows in KM3NeT
● Problem: feature design is not easy and maybe we missed
some good features?
● idea: let an algorithm learn the features directly on low-level
simulations
hits
track/shower?
neural network
Random decision forest
![Page 20: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/20.jpg)
convolutional
neural networks
![Page 21: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/21.jpg)
21Machine Learning workflows in KM3NeT
Successful model architecture in image recognition:
Convolutional neural networks (CNNs)
Simplified working principle of a CNN
Convolutional networks
![Page 22: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/22.jpg)
22Machine Learning workflows in KM3NeT
Our data
How does our data look like?
→ spatial: 3D detector
→ temporal
→ pmt direction: 31 orientations per DOM
DOM
![Page 23: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/23.jpg)
23Machine Learning workflows in KM3NeT
Our data: X-Y-plane
• line anchors in x-y plane are not on rectangular grid
• apply grid with <1 anchor per bin (assuming static detector)
→ 11 bins in X, 13 bins in Y (ORCA115)
JINST 15 P10005 (2020)
![Page 24: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/24.jpg)
24Machine Learning workflows in KM3NeT
Our data: Z-plane
• each line has 18 DOMs
→ similar heights for all lines
→ can be easily binned (assuming static detector)
18 bins
![Page 25: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/25.jpg)
25Machine Learning workflows in KM3NeT
Our data: time
• time coordinate is unbounded and continuous
→only use hits in a time window (e.g. 750 ns)
→ choose time resolution (e.g. 7.5 ns/bin)
• time resolution limited by hardware
• choose e.g. 100 time bins
JINST 15 P10005 (2020)
![Page 26: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/26.jpg)
26Machine Learning workflows in KM3NeT
Our data: pmt direction
DOM
31 pmts arranged on a 2 sphere
→ no spherical convolution in tensorflow,
so we use the color channel of
convolutions
→ instead of multiple colors, we supply
multiple pmt directions!
![Page 27: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/27.jpg)
27Machine Learning workflows in KM3NeT
2d plot, other
dimensions not shown
Input for convolutional networks
![Page 28: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/28.jpg)
28Machine Learning workflows in KM3NeT
Input XYZ-T
CNN
Input XYZ-P
OutputInput XYZ-T
Input XYZ-P
11 x 13 x 18 x (100+31)
● In total, we end up with 5d data (x, y, z, t, pmt)
● But tensorflow only supports up to 4D input to convolutions!
→ Solution:
● Stack two projections of the event: xyz-pmt and xyz-t
● use color channel of convolution for stacked dimension
Input for convolutional networks
softmax / cat. cross.
![Page 29: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/29.jpg)
29Machine Learning workflows in KM3NeT
Event topology classification
Neural networkvs
random forest
● Separability between track and shower
JINST 15 P10005 (2020)
![Page 30: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/30.jpg)
30Machine Learning workflows in KM3NeT
Background classification
Neural networkvs
random forest
● separate atmospheric muons and
neutrinos
JINST 15 P10005 (2020)
![Page 31: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/31.jpg)
31Machine Learning workflows in KM3NeT
● Convolutional networks on our data have various issues:
● no 5D convolution
● xyz positions need to be binned (problem for non-static detector)
● fixed time window with limited resolution
Graph networks
Idea: Use a network architecture that operates on graphs
![Page 32: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/32.jpg)
graph convolutional
neural networks
![Page 33: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/33.jpg)
33Machine Learning workflows in KM3NeT
Input for graph networks
![Page 34: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/34.jpg)
34Machine Learning workflows in KM3NeT
Edge convolution
each hit is a node 𝒙𝒊
= (𝑥, 𝑦, 𝑧, 𝑡, 𝑝𝑚𝑡) of the hit
2 nodes are connected with edge 𝒆𝒊𝒋
= (𝒙𝒊, 𝒙𝒊 − 𝒙𝒋) of the hits i,j
https://arxiv.org/abs/1902.08570
![Page 35: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/35.jpg)
35Machine Learning workflows in KM3NeT
Edge convolution
• define a multi-layer perceptron and
convolve over all edges
➢ produces an update 𝒖𝒊𝒋 from each edge
Then:
![Page 36: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/36.jpg)
36Machine Learning workflows in KM3NeT
Edge convolution
• define a multi-layer perceptron and
convolve over all edges
• Update central node 𝒙𝒊 with averaged
updates from k nearest neighbours:
𝒙𝒊 → 𝒙𝒊 + 𝒖𝒊𝒋 𝒋
➢ produces an update 𝒖𝒊𝒋 from each edge
Then:
![Page 37: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/37.jpg)
37Machine Learning workflows in KM3NeT
● Convolutional networks on our data have various issues:
● no 5D convolution
● xyz positions need to be binned (problem for non-static detector)
● fixed time window with limited resolution
Graph networks
Idea: Use a network architecture that operates on graphs
Fixed, can use n-D convolution
Fixed, unlimited resolution/time window
Fixed, no spatial binning necessary
![Page 38: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/38.jpg)
38Machine Learning workflows in KM3NeT
How good is the EdgeConv compared to convolutions?
→ faster, fewer parameters (→ less overfitting)
Convolution Graph
train time / epoch 8.3h 2.0h
free parameters 8.4m 370k
Graph networks
![Page 39: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/39.jpg)
39Machine Learning workflows in KM3NeT
▪ goal: reconstruct direction of atmospheric muons
Convolution Graph
0.0354 0.0349
Best validation loss
Graph networks: direction
loss (here: mean absolute error) is used to
judge performance of the reconstruction – the
lower the better!
(mean absolute error)
![Page 40: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/40.jpg)
40Machine Learning workflows in KM3NeT
▪ goal: reconstruct number of atmospheric muons in an event
Convolution Graph
0.389 0.361
Best validation loss:
Graph networks: multiplicity
(categorical cross-entropy)
![Page 41: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/41.jpg)
41Machine Learning workflows in KM3NeT
▪ goal: reconstruct distance between atmospheric muons
Convolution Graph
-1.911 -2.156
Best validation loss:
Graph networks: muon distance
(negative log-likelihood)
![Page 42: Machine Learning workflows in KM3NeT](https://reader033.fdocuments.in/reader033/viewer/2022052905/628ffddf35a53a1bc34002ba/html5/thumbnails/42.jpg)
42Machine Learning workflows in KM3NeT
Summary
▪ Machine Learning is an important tool for event reconstruction in KM3NeT
▪ allows to solve otherwise difficult to tackle problems
▪ workflows are improved continuously and adapted to our data