LIGO Detector Characterization in the Second and
Third Observing Runs
D Davis
1
, J S Areeda
2
, B K Berger
3
, R Bruntz
4
, A Effler
5
,
R C Essick
6
, R P Fisher
4
, P Godwin
7
, E Goetz
8
,
9
,
A F Helmling-Cornell
10
, B Hughey
11
, E Katsavounidis
12
,
A P Lundgren
13
, D M Macleod
14
, Z M ́arka
15
,
T J Massinger
12
, A Matas
16
, J McIver
17
, G Mo
12
,
K Mogushi
9
, P Nguyen
10
, L K Nuttall
13
, R M S Schofield
10
,
D H Shoemaker
12
, S Soni
8
, A L Stuver
18
, A L Urban
8
,
G Valdes
8
, M Walker
4
, R Abbott
1
, C Adams
5
,
R X Adhikari
1
, A Ananyeva
1
, S Appert
1
, K Arai
1
,
Y Asali
15
, S M Aston
5
, C Austin
8
, A M Baer
4
, M Ball
10
,
S W Ballmer
19
, S Banagiri
20
, D Barker
21
, C Barschaw
22
,
L Barsotti
12
, J Bartlett
21
, J Betzwieser
5
, R Beda
17
,
D Bhattacharjee
9
, J Bidler
2
, G Billingsley
1
, S Biscans
12
,
1
,
C D Blair
5
, R M Blair
21
, N Bode
23
,
24
, P Booker
23
,
24
,
R Bork
1
, A Bramley
5
, A F Brooks
1
, D D Brown
25
,
A Buikema
12
, C Cahillane
1
, T A Callister
26
,
27
, G Caneva
Santoro
28
,
29
, K C Cannon
30
, J Carlin
31
, K Chandra
32
,
X Chen
33
, N Christensen
34
, A A Ciobanu
25
, F Clara
21
,
C M Compton
21
, S J Cooper
35
, K R Corley
15
,
M W Coughlin
20
, S T Countryman
15
, P B Covas
36
,
D C Coyne
1
, S G Crowder
37
, T Dal Canton
16
,
38
, B Danila
39
,
L E H Datrier
40
, G S Davies
41
,
13
, T Dent
41
, N A Didio
19
,
C Di Fronzo
35
, K L Dooley
14
,
42
, J C Driggers
21
, P Dupej
40
,
S E Dwyer
21
, T Etzel
1
, M Evans
12
, T M Evans
5
,
S Fairhurst
14
, J Feicht
1
, A Fernandez-Galiana
12
, R Frey
10
,
P Fritschel
12
, V V Frolov
5
, P Fulda
43
, M Fyffe
5
,
B U Gadre
16
, J A Giaime
8
,
5
, K D Giardina
5
, G Gonz ́alez
8
,
S Gras
12
, C Gray
21
, R Gray
40
, A C Green
43
, A Gupta
1
,
E K Gustafson
1
, R Gustafson
44
, J Hanks
21
, J Hanson
5
,
T Hardwick
8
, I W Harry
13
, R K Hasskew
5
, M C Heintze
5
,
J Heinzel
34
, N A Holland
45
, I J Hollows
46
, C G Hoy
14
,
S Hughey
22
, S J Jadhav
47
, K Janssens
48
, G Johns,
4
,
J D Jones
21
, S Kandhasamy
49
, S Karki
10
, M Kasprzack
1
,
K Kawabe
21
, D Keitel
36
, N Kijbunchoo
45
, Y M Kim
50
,
P J King
21
, J S Kissel
21
, S Kulkarni
42
, Rahul Kumar
21
,
M Landry
21
, B B Lane
12
, B Lantz
3
, M Laxen
5
,
Y K Lecoeuche
21
, J Leviton
44
, J Liu
23
,
24
, M Lormand
5
,
R Macas
14
, A Macedo
2
, M MacInnis
12
, V Mandic
20
,
G L Mansell
21
,
12
, S M ́arka
15
, B Martinez
51
, K Martinovic
52
,
D V Martynov
35
, K Mason
12
, F Matichard
1
,
12
,
arXiv:2101.11673v2 [astro-ph.IM] 29 Jan 2021
2
N Mavalvala
12
, R McCarthy
21
, D E McClelland
45
,
S McCormick
5
, L McCuller
12
, C McIsaac
13
, T McRae
45
,
G Mendell
21
, K Merfeld
10
, E L Merilh
21
, P M Meyers
31
,
F Meylahn
23
,
24
, I Michaloliakos
43
, H Middleton
31
,
J C Mills
14
, T Mistry
46
, R Mittleman
12
, G Moreno
21
,
C M Mow-Lowry
35
, S Mozzon
13
, L Mueller
34
,
N Mukund
23
,
24
, A Mullavey
5
, J Muth
11
, T J N Nelson
5
,
A Neunzert
22
, S Nichols
8
, E Nitoglia
28
,
29
, J Oberling
21
,
J J Oh
53
, S H Oh
53
, Richard J Oram
5
, R G Ormiston
20
,
N Ormsby
4
, C Osthelder
1
, D J Ottaway
25
, H Overmier
5
,
A Pai
32
, J R Palamos
10
, F Pannarale
28
,
29
, W Parker
5
,
54
,
O Patane
2
, M Patel
4
, E Payne
55
, A Pele
5
, R Penhorwood
44
,
C J Perez
21
, K S Phukon
56
,
57
,
49
, M Pillas
38
, M Pirello
21
,
H Radkins
21
, K E Ramirez
51
, J W Richardson
1
, K Riles
44
,
K Rink
17
, N A Robertson
1
,
40
, J G Rollins
1
, C L Romel
21
,
J H Romie
5
, M P Ross
58
, K Ryan
21
, T Sadecki
21
,
M Sakellariadou
52
, E J Sanchez
1
, L E Sanchez
1
, L Sandles
14
,
T R Saravanan
49
, R L Savage
21
, D Schaetzl
1
, R Schnabel
59
,
E Schwartz
5
, D Sellers
5
, T Shaffer
21
, D Sigg
21
,
A M Sintes
36
, B J J Slagmolen
45
, J R Smith
2
, K Soni
49
,
B Sorazu
40
, A P Spencer
40
, K A Strain
40
, D Strom
11
,
L Sun
1
, M J Szczepa ́nczyk
43
, J Tasson
34
, R Tenorio
36
,
M Thomas
5
, P Thomas
21
, K A Thorne
5
, K Toland
40
,
C I Torrie
1
, A Tran
37
, G Traylor
5
, M Trevor
60
, M Tse
12
,
G Vajente
1
, N van Remortel
48
, D C Vander-Hyde
19
,
A Vargas
31
, J Veitch
40
, P J Veitch
25
, K Venkateswara
58
,
G Venugopalan
1
, A D Viets
61
, V Villa-Ortega
41
, T Vo
19
,
C Vorvick
21
, M Wade
62
, G S Wallace
63
, R L Ward
45
,
J Warner
21
, B Weaver
21
, A J Weinstein
1
, R Weiss
12
,
K Wette
45
, D D White
2
, L V White
19
, C Whittle
12
,
A R Williamson
13
, B Willke
24
,
23
, C C Wipf
1
, L Xiao
1
,
R Xu
37
, H Yamamoto
1
, Hang Yu
12
, Haocun Yu
12
,
L Zhang
1
, Y Zheng
9
, M E Zucker
12
,
1
, and J Zweizig
1
1
LIGO, California Institute of Technology, Pasadena, CA 91125, USA
2
California State University Fullerton, Fullerton, CA 92831, USA
3
Stanford University, Stanford, CA 94305, USA
4
Christopher Newport University, Newport News, VA 23606, USA
5
LIGO Livingston Observatory, Livingston, LA 70754, USA
6
University of Chicago, Chicago, IL 60637, USA
7
The Pennsylvania State University, University Park, PA 16802, USA
8
Louisiana State University, Baton Rouge, LA 70803, USA
9
Missouri University of Science and Technology, Rolla, MO 65409, USA
10
University of Oregon, Eugene, OR 97403, USA
11
Embry-Riddle Aeronautical University, Prescott, AZ 86301, USA
12
LIGO, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
13
University of Portsmouth, Portsmouth, PO1 3FX, UK
14
Cardiff University, Cardiff CF24 3AA, UK
15
Columbia University, New York, NY 10027, USA
16
Max Planck Institute for Gravitational Physics (Albert Einstein Institute),
D-14476 Potsdam, Germany
17
University of British Columbia, Vancouver, BC V6T 1Z4, Canada
18
Villanova University, 800 Lancaster Ave, Villanova, PA 19085, USA
3
19
Syracuse University, Syracuse, NY 13244, USA
20
University of Minnesota, Minneapolis, MN 55455, USA
21
LIGO Hanford Observatory, Richland, WA 99352, USA
22
University of Washington Bothell, Bothell, WA 98011, USA
23
Max Planck Institute for Gravitational Physics (Albert Einstein Institute),
D-30167 Hannover, Germany
24
Leibniz Universit ̈at Hannover, D-30167 Hannover, Germany
25
OzGrav, University of Adelaide, Adelaide, South Australia 5005, Australia
26
Stony Brook University, Stony Brook, NY 11794, USA
27
Center for Computational Astrophysics, Flatiron Institute, New York, NY
10010, USA
28
Universit`a di Roma “La Sapienza”, I-00185 Roma, Italy
29
INFN, Sezione di Roma, I-00185 Roma, Italy
30
RESCEU, University of Tokyo, Tokyo, 113-0033, Japan.
31
OzGrav, University of Melbourne, Parkville, Victoria 3010, Australia
32
Indian Institute of Technology Bombay, Powai, Mumbai 400 076, India
33
OzGrav, University of Western Australia, Crawley, Western Australia 6009,
Australia
34
Carleton College, Northfield, MN 55057, USA
35
University of Birmingham, Birmingham B15 2TT, UK
36
Universitat de les Illes Balears, IAC3—IEEC, E-07122 Palma de Mallorca,
Spain
37
Bellevue College, Bellevue, WA 98007, USA
38
Universit ́e Paris-Saclay, CNRS/IN2P3, IJCLab, 91405 Orsay, France
39
University of Szeged, D ́om t ́er 9, Szeged 6720, Hungary
40
SUPA, University of Glasgow, Glasgow G12 8QQ, UK
41
IGFAE, Campus Sur, Universidade de Santiago de Compostela, 15782 Spain
42
The University of Mississippi, University, MS 38677, USA
43
University of Florida, Gainesville, FL 32611, USA
44
University of Michigan, Ann Arbor, MI 48109, USA
45
OzGrav, Australian National University, Canberra, Australian Capital
Territory 0200, Australia
46
The University of Sheffield, Sheffield S10 2TN, UK
47
Directorate of Construction, Services & Estate Management, Mumbai 400094
India
48
Universiteit Antwerpen, Prinsstraat 13, 2000 Antwerpen, Belgium
49
Inter-University Centre for Astronomy and Astrophysics, Pune 411007, India
50
Ulsan National Institute of Science and Technology, Ulsan 44919, South Korea
51
The University of Texas Rio Grande Valley, Brownsville, TX 78520, USA
52
King’s College London, University of London, London WC2R 2LS, United
Kingdom
53
National Institute for Mathematical Sciences, Daejeon 34047, South Korea
54
Southern University and A&M College, Baton Rouge, LA 70813, USA
55
OzGrav, School of Physics & Astronomy, Monash University, Clayton 3800,
Victoria, Australia
56
Nikhef, Science Park 105, 1098 XG Amsterdam, Netherlands
57
Institute for High-Energy Physics, University of Amsterdam, Science Park
904, 1098 XH Amsterdam, Netherlands
58
University of Washington, Seattle, WA 98195, USA
59
Universit ̈at Hamburg, D-22761 Hamburg, Germany
60
University of Maryland, College Park, MD 20742, USA
61
Concordia University Wisconsin, 2800 N Lake Shore Dr, Mequon, WI 53097,
USA
62
Kenyon College, Gambier, OH 43022, USA
63
SUPA, University of Strathclyde, Glasgow G1 1XQ, United Kingdom
Abstract.
The characterization of the Advanced LIGO detectors in the second
and third observing runs has increased the sensitivity of the instruments, allowing
for a higher number of detectable gravitational-wave signals, and provided
confirmation of all observed gravitational-wave events. In this work, we present
the methods used to characterize the LIGO detectors and curate the publicly
4
available datasets, including the LIGO strain data and data quality products. We
describe the essential role of these datasets in LIGO-Virgo Collaboration analyses
of gravitational-waves from both transient and persistent sources and include
details on the provenance of these datasets in order to support analyses of LIGO
data by the broader community. Finally, we explain anticipated changes in the
role of detector characterization and current efforts to prepare for the high rate
of gravitational-wave alerts and events in future observing runs.
5
1. Introduction
The Laser Interferometer Gravitational-wave Observatory (LIGO) [1] and Virgo [2]
are the most sensitive facilities for the direct detection of gravitational-waves (GWs).
They have been observing the gravitational way sky in their advanced configuration
since 2015 and in a total three observing runs so far. Characterization of the LIGO
detectors enabled and enhanced the discoveries reported by LIGO-Virgo in their
second observing run (O2) and third observing run (O3). The two LIGO detectors
participated in O2 from November 30, 2016 to August 25, 2017, and the Virgo detector
joined for the last 25 days of the run. All three LIGO and Virgo detectors took data
during O3, from April 1, 2019 to March 27, 2020. The LIGO-Virgo Collaboration
has since reported the confident detection of gravitational wave signals from seven
black hole mergers and one binary neutron star merger during O2 in GWTC-1 [3] and
39 detections of black hole and neutron star mergers during the first half of O3 in
GWTC-2 [4].
The two US-based LIGO detectors are dual-recycled Michelson interferometers
with 4 km Fabry-Perot arm cavities. The LIGO detectors are designed to sense
extremely small fluctuations in spacetime induced by passing gravitational waves [1].
LIGO Hanford (LHO) is located in Hanford, Washington, and LIGO Livingston
(LLO) is located in Livingston, Louisiana. During O2 and O3, some differences
in configuration between the LHO and LLO instruments resulted in differences in
technical noise sources contributing to effective gravitational wave strain noise between
the two detectors, as reported in [5, 6].
LIGO detector data is a gravitational-wave strain time series (referred to as
the
GW strain
) that is rich with noise artifacts. Often the noise is dominated by
fundamentally limiting noise sources [1], causing it to appear Gaussian and stationary
over limited time scales and frequency ranges. The sensitivity of the detectors as
measured by the amplitude of these noise sources, combined with the coincident
uptime with multiple detectors observing, are key metrics of detector performance.
However, LIGO data also contains a high rate of transient noise artifacts, or
glitches
,
that contribute to the noise background of searches for gravitational waves by
mimicking the behavior of true astrophysical signals. Glitches can also overlap with
signals, as reported in [7], and confuse source property estimation of even confidently
detected signals unless properly mitigated [8, 9, 10]. LIGO data also contains strong
nearly sinusoidal features, or
lines
, that inhibit searches for long duration sources of
gravitational waves, as described in [11]. Additionally, LIGO detector data exhibits
slow changes to the characteristics of the noise due to complex interactions between
the detectors and their local environment.
In order to address these features of the data that differ from the output of an
idealized gravitational-wave interferometer, the LIGO detectors and data are closely
monitored before and during observing runs using a large number of additional data
streams (referred to as
auxiliary channels
), that include sensors of the environment
surrounding the detectors and measurements of the detector control systems. These
efforts to understand and mitigate these sources of noise, both in the instrument
and the data are collectively referred to as “detector characterization”. Detector
characterization is an essential component of improving the performance of the LIGO
detectors and the detection of gravitational wave events [12, 13].
LIGO data is publicly distributed via the Gravitational-wave Open Science Center
(GWOSC) [14]. Currently available data includes GW strain data during periods the
6
individual detectors were observing in the first two observing runs and data quality
information used in LIGO analyses [15, 16]. LIGO data from the third observing
run is planned to be released in six month periods, 18 months after the start of
each observing period [17]. In addition to these bulk data releases, data nearby all
detected gravitational-wave events is released via GWOSC at the time of publication.
Data from a subset of auxiliary channels is currently available for a three hour period
around a single event [18].
In this paper, we report the results of detector characterization methods applied to
LIGO detector data from O2 and O3 to improve the performance of the detectors and
astrophysical analyses. In section 2 we summarize the LIGO O2 and O3 data sets as
reported in [3, 4]. In section 3 we describe major tools and infrastructure employed for
LIGO detector characterization during these observing runs. In section 4 we outline
work that improved the performance of the LIGO detectors by characterizing and
mitigating sources of instrumental noise. In section 5 we summarize the methodology
of LIGO data quality products employed by transient gravitational wave searches using
O2 and O3 data as well as methods and procedures applied to LIGO detector data to
validate transient event candidates. In section 6 we describe data quality investigations
and products used by searches for gravitational waves from persistent sources. We
conclude in section 7 with an overview of future work, including automation efforts
designed to cope with the significantly higher sensitivity and expected event rate
during future observing runs.
2. The O2 and O3 data sets
The O2 period spanned 268 calendar days, with the LIGO detectors participating
for the entire period. Virgo, however, joined for the last 25 days. There were two
scheduled breaks over the observing run; the 2016 end-of-year holidays and a few
weeks in May 2017, which was used to make improvements to each of the LIGO
detectors.
One way in which we measure sensitivity is by the binary neutron star inspiral
range; this range is the distance at which a gravitational-wave signal from a the
merger of two 1
.
4
M
neutron stars would be detected above signal-to-noise ratio
(SNR) 8, averaged over all possible sky locations and inclinations without considering
cosmological corrections. The LLO detector started O2 observing around 80 Mpc,
and became steadily more sensitive as O2 progressed, reaching 100 Mpc.
The
LHO detector’s sensitivity was around 75 Mpc at the start of the observing run.
It however, suffered a sudden drop in sensitivity on 6th July 2017 due to a 5.8
magnitude earthquake in Montana, finishing the run around 65 Mpc. Virgo held
a steady sensitivity around 25 Mpc for its 25-day observing period. This information
is illustrated in figure 1.
The O3 observing run was split into two periods, separated by the month of
October 2019 to make stability improvements to all three detectors. The first half
of the third observing run (O3a) lasted for 183 days with the LLO, LHO and Virgo
detectors having a median range of 135 Mpc, 108 Mpc and 45 Mpc respectively.
Due to the improvements made to the interferometers [4] between O2 and O3, the
sensitivity of the detectors increased by a factor of 1.53 for LLO, 1.64 for LHO and
1.73 for Virgo. During second half of the third observing run (O3b) the sensitivity of
the detectors were similar to O3a, with LLO, LHO and Virgo each having a median
range of 131 Mpc, 113 Mpc and 50 Mpc. O3b lasted 147 days, some 34 days less than
7
Duty Cycle
Detector
O2
O3a
O3b
O3
LHO
65%
71%
79%
75%
LLO
62%
76%
79%
77%
Virgo
85%
76%
76%
76%
LHO+LLO
46%
59%
67%
62%
LHO+LLO+Virgo
63%
44%
51%
47%
Table 1: The duty cycle (round to the nearest integer) of each of the detectors, LIGO
Hanford (LHO), LIGO Livingston (LLO) and Virgo, and combinations, over
the second (O2) and third observing run (O3).
The O3 numbers are a
combination of O3a and O3b.
was originally intended. If we instead consider the sensitivity of the detectors to the
inspial of two black holes each with a mass of 30
M
, the ranges become approximately
1425 Mpc, 1150 Mpc and 525 Mpc for LLO, LHO and Virgo respectively, throughout
O3.
Figure 1 shows the typical amplitude spectral density of the strain noise for each
detector over O2 and O3. The duty cycle of each detector defines the amount of sci-
ence quality data taken over a period of time. There are a number of factors which
affect the duty cycle, such as the environment (e.g., weather), detector hardware (e.g.,
malfunctioning instrument components) and periods of commissioning. Table 1 high-
lights the duty cycle of each of the detectors in O2 and O3. We also give the coincident
duty cycle of the LIGO detectors, as well as the triple coincident time. There is a
marked improvement in the stability of the LIGO detectors between O2 and O3, with
coincident science quality time increasing by some 16%. Although the Virgo duty cy-
cle appears to decrease between observing runs, it should be highlighted that the O2
duty cycle includes livetime that is about 13 times less than in O3. As well, the 25-day
O2 time that Virgo was observing for included optimal environmental conditions.
In both observing runs, we used auxiliary channels that recorded the source of the
instrumental noise (referred to as a “witness”) to measure sources of noise that limit
detector sensitivity. Using these measurements, we were able to linearly subtract this
noise from the data. During O2 a pipeline was developed to do this subtraction, which
is easily adaptable to target new sources of noise as they arise [5, 19]. For both LIGO
detectors this was used to target narrow line features, such as the calibration lines and
60 Hz and its harmonic frequencies. At LHO, however, there was an additional source
of broadband noise known as jitter noise. This form of noise was related to the jitter
of the pre-stabilized laser beam in angle and size. This was only present at LHO due
to different configurations between the two LIGO detectors. By subtracting this form
of noise below 1000 Hz, LHO saw an average increase in range, over O2, of 20% [19].
The LLO detector saw no appreciable increase in its range.
In O3 issues of jitter noise had been resolved, and so the same level of data
cleaning was not necessary. The removal of the calibration lines and noise from their
harmonics were subtracted from data as part of the calibration procedure. For a
subset of gravitational wave events detected during O3, additional data cleaning was
performed to remove noise contributions due to non-stationary couplings of the power
8
0
10
20
30
0
50
100
150
BNS Range [Mpc]
LIGO Hanford
LIGO Livingston
Virgo
130
140
150
160
170
Weeks since O2 start
10
1
10
2
10
3
Frequency [Hz]
10
−
24
10
−
23
10
−
22
10
−
21
10
−
20
10
−
19
Strain [1
/
√
Hz]
O2 : LIGO Hanford
O3 : LIGO Hanford
O2 : LIGO Livingston
O3 : LIGO Livingston
O2 : Virgo
O3 : Virgo
Figure 1: Top: Binary Neutron Star (BNS) range evolution of the LIGO and Virgo
detectors from the start of O2 in November 2017 to the end of O3 in
March 2020. The broken axes remove the time between each observing run.
Bottom: Representative amplitude spectral density of the three detectors’
strain sensitivity in each observing run. The O3 spectra shown are taken
from O3a.
mains [20].
As the interferometers are upgraded and improved, and more hardware goes
into the interferometers, inevitably, there are more data quality issues that arise.
9
For instance, O3 saw the installation of squeezed light sources [21, 22]. This is an
additional system that could and did introduce additional noise into the O3 data.
As the detectors become more stable, not only does their duty cycle improve, the
increased observing time allows for more data quality issues to occur. These data
quality issues are discussed in detail in the remainder of this paper.
3. Computing and Software
Because ground-based gravitational-wave detectors are subject to a wide range of
environmental noise source and continual upgrades during an observing run, and
because new technologies periodically emerge to improve sensitivity across the full
frequency range, the detectors themselves are continually evolving. This presents an
endless challenge to any effort to characterize noise in the detectors, as the source,
shape, rate, and intensity of various noise sources is constantly changing. In this
section we will outline the multiple computational solutions which have emerged
to help combat this problem, focusing on the types of analyses that each software
application is suited to. In doing so, we will build important context for the methods
and results presented later in the paper. This section is not meant as an exhaustive
list of all analysis tools that are used in detector characterization studies, but it serves
to give a broad example and context for results discussed in this article.
3.1. Signal processing tools
A number of open source computing projects are developed and maintained to enable
data analysis for LIGO detector characterization. These tools are published through
widely used version control platforms and delivered to users through the International
Gravitational-wave Network (IGWN) Conda Distribution [23].
Unless otherwise
noted, they are written entirely in Python [24] and are available under terms of the
GNU General Public License, version 3.0.0.
These signal processing tools are designed to both process raw timeseries that are
generated by the wide variety of data streams at each observatory, as well as other
pre-processed data. One of the main types of pre-processed data types that these
tools are designed to ingest are “triggers” created by event trigger generators (ETGs).
A wide variety of ETGs exist, but in general, are designed to find excess power in
data streams. These excess power bursts are considered triggers. While some ETGs
are designed to identify generic bursts of excess power using wavelets, other ETGs
use waveform templates from general relativity to identify triggers that are consistent
with a particular gravitational-wave source.
3.1.1. GWpy
The central signal processing and data visualization engine used to
prepare most figures in this paper is
GWpy
[25, 26], a Python package for studying
data from gravitational-wave detectors. This package is designed with an extensive set
of features for manipulating data in both the time and frequency domain, including:
(i) Native memory-optimized Python classes for
TimeSeries
and
FrequencySeries
objects
(ii) Robust data input/output capabilities, including support for multiple file formats
as well as time optimization through multithreading
10
(iii) Custom filtering applications, digital filter designers, and convolution algorithms
tailored for IGWN data
(iv) An implementation of both the fast Fourier transform and the multi-Q transform
(see section 3.1.3) for timeseries data
(v) A
Table
class primarily designed for analyzing the output of various ETGs
(vi) Publication-quality visualization methods that are fine-tuned for every data
product while remaining highly customizable
While
GWpy
is aimed at individual users and is relatively general-purpose, a
number of other packages with narrower scope are also derived from it, as described
below.
3.1.2. GW-DetChar
An extension of
GWpy
with specific applications to IGWN
(especially LIGO) detector characterization tasks is available in the
GW-DetChar
software package [27]. This codebase contains a number of user modules and scripts
that are able to identify and analyze known classes of glitches, such as optical
scattering [28], as well as more general noise hunting algorithms such as Lasso
regression [29]. For each tool in the package, the primary data product is a single
webpage with responsive design features [30] which can be used to record and easily
interpret results (see section 3.2.2).
3.1.3.
Omega scans
A particular data visualization submodule within
GW-
DetChar
is the
gwdetchar-omega
command-line tool, so named because it is a
Python implementation of a legacy unmodeled transient search pipeline called Omega
[31, 32, 33]. In a detector characterization context this tool is used to identify and
visualize the time-frequency morphology of various sources of transient noise. Its
primary data product is referred to as an omega scan, which consists of a raw multi-Q
transform optimized over the quality factor. The optimized raw constant-Q transform
is then interpolated, providing a qualitative high-resolution image of signal energy as
a function of time and frequency (i.e., a spectrogram).
Through configuration files, users have the ability to analyze an arbitrary number
of data streams over the same time range, which makes the omega scan a powerful
tool in tracing the propagation of a glitch throughout interferometer subsystems.
To assist with this,
gwdetchar-omega
can optionally cross-correlate every successive
independent data stream with the signal, then display a tabulated ranking of the
most highly correlated channels. Alternatively, the omega scan can be used to process
GW strain streams from an arbitrary number of interferometers for a quick visual
comparison of the signal morphology in each stream over a fixed time interval.
3.1.4.
Omicron
The primary ETG for detector characterization studies is an
unmodeled transient detection pipeline called Omicron [34]. The Omicron pipeline
broadly performs a multi-Q transform given some data stream, then searches for
significant clusters of tiles in time-frequency space, optimizing over the quality
factor. For each LIGO detector, Omicron is run on the GW strain channel and a
collection of some 800-900 separate channels representing interferometer subsystems,
with triggers stored in a central location from their on-site computing clusters. To
ensure stability of trigger production, the workflow is managed by a Python package
called
pyomicron
[35] and most channels have triggers available with modest 1 hour
latency.
11
3.1.5. Hierarchical Veto
While omega scans can be used to identify correlations
between an arbitrary number of data streams, such analyses tend to be confined to a
narrow window of time (
∼
1 sec) to understand the origin of a specific transient glitch.
On the other hand, it is well worth understanding broader, longer-term correlations
that may exist over the course of hours or days and to assign a statistical signifcance
to the identified correlation. This is the purview of
HVeto
[36, 37], a companion to
GW-DetChar
that analyzes concurrent patterns between clusters of event triggers
above a fixed SNR threshold in multiple data streams.
HVeto
correlation searches are used to identify potentially statistically
significant coincidences between Omicron [34] triggers in the GW strain channel and
other auxiliary channels. The significance is calculated as the probability of the
number of observed coincidences divided by the number expected (See [36]).
HVeto
is used multiple times per day to correlate glitches with auxiliary channels that may
interfere with the identification of gravitational waves. When a significant association
is found,
HVeto
analyzes the effect of removing the time segments containing the
associated glitches from the analysis before proceeding to other data streams in a
hierarchical fashion. With each successive round of vetoes, a list of statistically
significant correlations emerges, ranked in descending order of significance. To better
understand the cause of the identified correlations, omega scans of a subset of the time
periods removed in each round are generated for visual inspection.
3.2. Web-based services
By contrast with the software packages described in section 3.1, the following services
are maintained as broad signal processing platforms primarily accessible to the end
user through the Internet, with application programming interfaces (APIs) available
on the command-line and through any computing environment that supports Python.
Like most LIGO and IGWN web-based services, they utilize Shibboleth Single Sign-
on [38] for user authentication to ensure the security of proprietary datasets.
3.2.1. DQSEGDB
Many detector characterization tasks and pipelines designed to
search for GW signals rely on data quality (DQ) “flags.” DQ flags store metadata
for measured or derived states within each interferometer, its subsystems, and various
components. DQ flags are used at all current GW observatories.
These DQ flags are stored in the IGWN Data Quality Segment Database
(DQSEGDB) [39]. For each flag, the database tracks spans of time (called
segments
)
over which the flag’s on/off truth value is known. In particular, each interferometer’s
“observing mode” flag indicates segments over which that interferometer was both
locked and taking science-quality data that is flagged by interferometer operators as
intended for GW searches. This flag is used by all downstream analysis pipelines to
distinguish spans of time that can and cannot be analyzed. Other flags can be used
to reject (or
veto
) otherwise usable segments in which the data stream contains well-
understood artifacts, such as glitches with a known cause, or planned injections of
artificial test signals.
3.2.2.
Detector characterization summary pages
For the convenience of LIGO
commissioners, detector engineers, and data analysts, an extensive suite of detector
characterization summary pages is provided [40] which offer automated daily analyses
of the primary GW strain data as well as sundry interferometer subsystems. These
12
pages are available to anyone with federated credentials on the
LIGO.ORG
domain, and
are batch-generated on dedicated hardware with modest (0.5-1 hour) latency. The
detector characterization summary pages are one of the main tools used to monitor
the performance of the LIGO interferometers and the data quality. The centralized
location of these automated analyses also allows for detailed follow up of any identified
issues.
The raw HTML is built programmatically through the
GWSumm
software
package [41], an extension of
GWpy
that also manages core signal processing,
while interactive webpage elements are implemented through JavaScript. The visual
layout of the front end is color-coded by interferometer, with responsive web design
accomplished through Bootstrap [42] and a custom extension thereof called GW-
Bootstrap [30].
‡
While the LIGO detector characterization summary pages are built on-the-fly
via Python code, they are also designed to be easily tunable through configuration
files.
Users have the freedom to register, design, and build a diverse array of
visualizations, ranging from simple timeseries tracks to more complicated time-
frequency spectrograms, with fine-grain control over all signal processing parameters.
Because the back end utilizes Asynchronous JavaScript and XML [44] web
development techniques, users also have the option to build their own custom images
and HTML on the server side, then load them remotely through the summary pages.
This workflow allows several third-party analyses to be hosted in one centralized
location, including automated data-quality products for persistent GW searches.
While the
GWSumm
software package is most heavily used by LIGO, it is
designed for use by the broader IGWN community. A suite of pages is currently
built using the same software for the KAGRA detector [45, 46], while independent
software provides a very similar service for the Virgo detector [47]. A less extensive
public-facing version of the summary pages, built with
GWSumm
and focusing only
on time segments and GW strain data, is also available [48].
3.2.3. LigoDV-web
The
LIGO DataViewer Web
service (LDVW) [49] is an online
data visualization platform providing direct interactive access to data recorded at the
LIGO Hanford and Livingston observatories and a subset of data from Virgo, KAGRA,
the GEO600 observatory in Hanover, Germany [50], and the smaller 40m prototype
interferometer in Pasadena, CA, USA [51]. This software instantaneously provides
users with custom visualizations of small data sets in a fast, secure, and reliable
manner and with minimal software, hardware, and training requirements. LDVW
adds a convenient online tool that allows the generation and sharing of custom data
visualizations to augment standardized analyses such as those on the Summary Pages.
It is often the most convenient way to access the large number of different data sources
at each site and generate large numbers of plots to address specific questions.
LDVW is implemented as a Java Enterprise application [52] with a proprietary
network protocol used for data access on the back end [53]. LIGO-Virgo-KAGRA
Collaboration members with proper credentials can request data to be displayed
in several formats from any Internet appliance that supports a modern browser
with JavaScript and minimal HTML5 support, particularly personal computers,
smartphones, and tablets. The primary signal processing and image rendering engine
‡
To keep the file directory structure clean, these packages are published through the Node.js Package
Manager (
npm
) [43] and supplied to the LIGO summary pages via content delivery networks.
13
for this service is
gwpy-plot
, a robust command-line interface for
GWpy
[25].
3.2.4. Data Quality Reports
In the context of detector characterization, a
Data
Quality Report
(DQR) [54] is an internal collection of convenient analysis routines
used to support and enable the vetting of GW event candidates. It is tightly integrated
with the
LIGO-Virgo Alert System
(LVAlert) [55] and
Gravitational-Wave
Candidate Event Database
(GraceDB) [56]. When an upstream search pipeline
identifies a potential GW signal, the event is recorded in GraceDB and the LVAlert
system broadcasts a notice to all subscribers, including the DQR architecture. When
the DQR receives an alert it triggers a series of analyses from three LIGO computing
clusters. Examples of included analyses are omega scans, statistical checks such as
HVeto
, and checks of known flags in DQSEGDB. The DQR infrastructure is modular,
allowing for additional tools to be added as desired.
Within minutes of the initial GW event candidate, the DQR architecture begins
to upload web-based reports and supporting data to GraceDB for internal review,
which then informs the decision to disseminate additional the Gamma-ray Coordinates
Network (GCN) Notices and Circulars [57] or to retract an announced candidate.
Additional details about the tools currently implemented in the
Data Quality
Report
and related event validation procedures are described in section 5.5.
3.2.5. Spectral artifact tools for persistent GW searches
Several different tools have
been developed to aid in finding narrow, persistent spectral artifacts in GW detector
data [11]. These tools build amplitude spectral density plots using fast Fourier
transforms (FFTs) that are 1800 s, or longer, over time periods of 1-day up to an
entire observing run. Since the coherent baseline is much longer than other figures-of-
merit, and averaged across epochs, it allows for understanding the narrow, persistent
spectral artifacts that corrupt searches for continuous GWs from spinning neutron
stars.
One of these tools, known as
Fscan
, runs automatically each day and generates
1800-s-long FFTs for the low-latency GW strain channel and a subset of auxiliary
detector channels and physical environment monitoring channels. Various figures of
merit can be derived from the FFT data computed from the primary GW strain
channel and the subset of additional channels. This enables more regular monitoring
of the behavior of spectral artifacts.
For example, normalized, day-, week-, and month-long averaged amplitude
spectral densities (ASDs) are computed from the FFT data as well as coherence
between the GW strain channel and the subset of additional channels. Correlations of
spectral artifacts in ASDs of different channels can then be identified, or via coherence,
as possible non-astrophysical causes of spectral artifacts. Coherence is a useful figure of
merit to reject spurious coincidence of spectral artifacts that are not actually correlated
and to identify potential coupling mechanisms of non-astrophysical noise into GW
data.
Specifically, coherence between two channels
d
1
(
t
) and
d
2
(
t
) is defined as [11]
Γ(
f
) =
〈|
̃
d
?
1
(
f
)
̃
d
2
(
f
)
|
2
〉
〈|
̃
d
1
(
f
)
|
2
〉〈|
̃
d
2
(
f
)
|
2
〉
,
(1)
where
̃
d
i
(
f
) (
i
= 1
,
2) is the Fourier transform of the time series data
d
i
(
t
),
?
denotes
complex conjugation, and the average
〈·〉
refers to an average over
N
segments. For
14
Gaussian, uncorrelated noise, the expected distribution of the coherence is given by
p
(Γ)
∝
e
−
Γ
N
,
(2)
where
N
is the number of segments used for averaging.
Other examples of figures of merit include:
FineTooth
, a comb identification and
tracking tool;
NoEMi
, a line monitoring and database tool; a coherence tool database
that enables efficient look-up of coherences in different time- and frequency-intervals;
and studies that fold time-domain data at periodic intervals to check for periodic
elevated noise. Additional figures of merit are under development for future use to aid
understanding of spectral artifacts in GW detector data.
4. Instrumental Investigations
In order to maximize the opportunities for the discovery of astrophysical gravitational-
wave signals, it is essential to understand the instrumental and environmental noise
that can mimic or obscure such signals. Recognition of potential noise couplings can
lead to detector hardware changes to reduce the rate of noise artifacts. In this section,
we first describe our approach to identifying instrumental noise and mitigating its
effects on astrophysical searches. Later we discuss the major types and sources of
transient noise and their impact on detector data quality.
4.1. Instrumental Investigation methods
4.1.1. Data Quality Monitoring
The LIGO instruments are delicate in the sense
that glitches or other manifestations of noise can appear in the GW strain channel.
During the observing period, a “shifter” assigned at each site conducts a week long
data quality shift. The objective of the DQ shifter is to monitor the behavior of the
instruments, note any changes, and to communicate them to the commissioners at
LHO and LLO and the members of the detector characterization group. The LIGO
summary pages [40] are the typical launching point for off-site DQ investigations of
instrument noise. These pages provide both an overview of the detector status and
the low-level information about specific detector subsystems. The summary pages
also display the results of analysis algorithms that identify or correlate noise in both
auxiliary sensors and gravitational-wave strain data. The computing infrastructure of
the summary pages is discussed in further detail in section 3.2.2.
Omega scans,
HVeto
, Lasso [29, 27] and Omicron are some of the most commonly
used analysis tools for identifying noise in the detector. Tracking down sources of
transient noise often begins with the output of Omicron (see section 3.1.4), which finds
short-duration bursts of noise. The resulting events, or “triggers”, can be plotted in
the time-frequency plane to visualize transient noise in both witness auxiliary sensor
and GW strain data. For each day, the Omicron triggers (glitches) for the GW
channel are shown on a time-frequency plot with the markers color-coded for SNR.
An increase in the number of high SNR glitches indicates a change in the instrument
or environment. If glitches persist at specific frequencies critical to data analysis,
they constitute a problem for event detection and parameter estimation. Therefore,
it is essential to eliminate the harmful influence of these glitches on the searches for
events. To recognize the potential sources of problematic glitches, we use
HVeto
(see
section 3.1.5) to identify witness auxiliary sensors whose bad behavior coincides with
the appearance of a subclass of glitches.
15
10
100
500
Strain Data
0
.
0
1
.
0
2
.
0
3
.
0
4
.
0
5
.
0
6
.
0
Time [hours]
10
100
500
Squeezer Data
0
5
10
0
1
2
Frequency [Hz]
Relative amplitude
Figure 2:
Top
: Spectrogram of the relative amplitude of gravitational-wave strain
timeseries (
h
(
t
)) during the first 6 hours (UTC) of 2019-04-26. Between about
3h and 6h, at frequencies of about 40 Hz to 300 Hz, peculiar line features are
visible. These are the wandering (in frequency) lines discussed in the text.
Bottom
: Squeezer data when the wandering line in
h
(
t
) is visible. Features
(in yellow) in this squeezer channel match those seen in
h
(
t
) in timing and
frequency. See [58] for details.
As an example of the value of DQ monitoring, we consider the “squeezer
wandering line” at LHO as described here [59, 60]. A line-like feature with changing
frequency was clearly visible in
h
(
t
) hourly histograms (see figure 2). This equally
spaced comb of lines will appear and disappear suddenly between 80 Hz and 140 Hz
at LHO. Similar wandering lines were noticed at LLO between 160 Hz and 200 Hz.
This feature was shown to be correlated in time and frequency with a strong feature
in several squeezer (SQZ) channels at both the detectors. The DQ issue was solved by
turning off the squeezer laser “noise eater” at LLO [58]. This was then implemented
successfully at LHO to cure the problem [61]. See [62, 63] for examples of additional
solved transient noise issues.
4.1.2. Physical Environment Monitoring
Environmental noise can affect a LIGO
detector by limiting its sensitivity to astrophysical GW signals and producing
transients in the strain data. Some environmental noise sources can potentially
be correlated between detector sites, making them particularly problematic for
astrophysical searches. It is, therefore, important to identify these sources and mitigate
their effects. The methodology and hardware for investigating environmental noise are
discussed in detail in [64, 65]. They typically consist of generating a “noise injection”
of known amplitude and frequency range and observing the detector’s response to
the signal. Sensors that monitor the physical environment are used to measure the
injection and estimate the noise source’s coupling to the detector. For example,
we estimate acoustic coupling by using accelerometers and microphones to measure
acoustic noise injections made from speakers. The sensors are much more sensitive to
environmental noise than the detector is, making them good witnesses of noise sources
that couple into the GW strain channel. Other injection methods include shaking