Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
https://ntrs.nasa.gov/search.jsp?R=20170012488 2019-08-30T16:40:42+00:00Z
A Search for Earth-size Planets
Overview
• What did it take to build the Kepler science pipeline?
• Major modifications to pipeline over lifetime
• High fidelity simulations• Commissioning, commissioning,
commissioning• High performance computing• Developing the TESS Science Pipeline• Communication• Summary
A Search for Earth-size Planets
A Search for Earth-size Planets
• Design started in earnest in 2004 with launch in March 2009 and operations through May 2013 and reprocessing through 2017
• A total of ~100 person years of effort went into the first complete version of the pipeline (from pixels to planets)
• The staffing was at ~20 individuals per year through 2016, tapering off thereafter (~280 FTEs over project lifetime)
• Build 5.0 was the launch-ready software release• There were 4 major builds thereafter, with substantive point
releases to mitigate issues subsequently identified in flight or full volume re-processing
• Build 9.0, 9.1, 9.2, 9.3 really represented at least two full builds of effort (issues identified in full re-processing and in completeness and reliability processing)
• Unexpected instrumental effects/stellar variability/hardware failures motivated significant software modifications on orbit
The Science Operations Center: What did it take?
A Search for Earth-size PlanetsScience Operations Center Architecture
A Search for Earth-size Planets
CALPixel level Calibration
PAPhotometric
Analysis
TPSTransiting Planet
Search
PDCPre-search Data
Conditioning
RawData
CorrectedLight Curves
Calibrated Pixels
Raw Light Curves & Centroids
DVData Validation
Diagnostic Metrics & Reports
Threshold Crossing Events(TCEs)
TCERTThreshold Crossing Event
Review Team+ Robovetting
Artificial Transit & BEB Injection Machine
Artificial Transits and Eclipses
Planet, or dud?Auto-Vetting
Applying machine learning to candidate
evaluation
>1,000,000 Lines of Code; 26 different Modules
Kepler’s Science Pipeline
A Search for Earth-size Planets
Every component of the science pipeline saw major evolution over mission Pixel level calibrations: • Updates based on actual electronics behavior• Flagging of electronic image artifacts causing false positivesIdentifying optimal apertures• Use of reconstructed pointing• Added ability to correct errors in Kepler Input catalogPhotometric analysis• Major improvements to identifying cosmic raysPre-search Data Conditioning• Development of Maximum a Posteriori approach• Addition of multi-scale analysis• Detection of Sudden Pixel Sensitivity DropoutsTransiting Planet Search• c2 vetoes addedData Validation• Difference image analysis• Ghost Diagnostic + other metrics
Major Modifications
A Search for Earth-size PlanetsShort Timescale Instrumental Errors
Kepler is sensitive to its thermal environment
Signature of a heater cycling on the reaction wheels 3/4
A Search for Earth-size PlanetsInstrumental Effects in Photometry
High pass filtered
A Search for Earth-size PlanetsCorrecting Systematic Errors
Original Flux
Systematic Error-Corrected Flux
We apply a Maximum A Posteriori approach as per Stumpe et al. 2014
A Search for Earth-size Planets
End-End Model (ETEM) drove design of SOC and testing of entire ground segment
Simulated data were so good that we didn’t need to update the compression tables after launch (the achieved compression (~4.5-5 bits per pixel) was within 0.1 bits of ideal performance
High Fidelity Simulations are Indispensable
0 5 10 15 20 25 30
4
5
6
7
8
9
10
Time, Days Since Start of Q3
En
tro
py,
Bits
Pe
r P
ixe
l
AchievedTheoretical
A Search for Earth-size Planets
Difference image analysis was key for Kepler for excluding false positives from background eclipsing binariesEspecially important for bright, saturated (bleeding) targets
Difference Image Analysis
6 Rappaport et al.
Fig. 5.— Direct images and di↵erence images for KIC 3542116 during each of the three major transit features for each quarter in whichthey occur. Top panels show the mean calibrated pixel values in the aperture masks returned by Kepler for Q10, Q12 and Q13, from leftto right. Bottom panels show the mean di↵erence between the calibrated pixels in transit-feature-wide segments on either side of eachtransit, and the pixel values during each transit feature. The source of the transit features will exhibit positive values in these di↵erenceimages. The colorbars indicate the pixel fluxes in units of e�/s. Note that the pixels exhibiting the strongest positive deviations in thebottom panels occur primarily at the ends of the saturated and bleeding columns and are approximately clustered about the location ofKIC 3542116, which is marked by a red star. KIC 3542117’s location is marked by a red circle in each panel. These di↵erence imagesindicate that the source of the transit features is co-located with KIC3542116 to within the resolution of the Kepler data.
so that dim images (and sometimes negative images) ofstars read out on the adjacent three CCD readout chan-nels are electronically superimposed on the image databeing read out by the fourth CCD channel (van Cleve &Caldwell 2016).We looked for stars located on the Kepler detectors
which might be the source of any video cross-talk sig-nals by inspecting the full frame images (FFIs) for thequarters during which the three most prominent transitsoccurred. There is a fairly bright, possibly saturated starnear the edge of the optimal aperture on output 3 on theCCD on which 3542116 is imaged (it’s on output 1 in allcases), but the crosstalk coe�cient is even smaller thanfor the other two outputs, �0.00001, so that a 50% deepeclipse on this other star would be attenuated to a valueof 5 ppm when its video ghost image is added to the di-rect image of KIC 3542116, assuming they are the samebrightness. Furthermore, since the coe�cient is negative,there would need to be a brightening event on the staron output 3 to cause a transit-like dip on output 1.Fortunately, the largest crosstalk coe�cient to the
CCD output that 3542116 finds itself on is +0.00029, sogiven that the signal we are looking at is ⇠0.1%, a con-taminating star would need to be at least 10⇥ brighterthan 3542116 to cause a problem. If there were, it wouldbe highly saturated and bleeding, which would make itdi�cult to square with the pixel-level analysis indicat-ing that the source is associated with the pixels under3542116, as the extent of the bleeding would be signifi-
cantly larger than for KIC 3542116.We also inspected the quality flags associated with the
flux and pixel time series for KIC 3542116 and find thatthe situation is nominal with flags for occasional eventssuch as cosmic rays and reaction wheel desaturations, butno flags for rolling band noise during the transit events.16
Finally, we considered ‘Sudden Pixel SensitivityDropouts’ (SPSDs) in the data, which are due to ra-diation damage from cosmic ray hits on the CCD, as apossible explanation for the dips in flux that we observe.However, the shape and behavior of such dips do notresemble what we see (Thompson et al. 2016b). In par-ticular, the SPSD events have drops that are essentiallyinstantaneous, and therefore are much shorter than the⇠20 and 8 long-cadence points on the ingresses that wesee in the deeper and more shallow dips, respectively.Moreover, the location of the SPSDs on the CCD chipwould not plausibly align with the source location andits bleed tracks for each and every one of the dips. Thus,we also discarded this idea as well.We take all these evaluations as strong evidence that
the dips we see are of astrophysical origin and that KIC3542116 is indeed the source of them. However, we can-not categorically rule out the possibility that the dips arecaused by some unknown peculiar type of stellar variabil-ity in KIC 3542116 itself. In spite of this caveat, weproceed under the assumption that the dips in flux are
16 See Thompson et al. (2016b) for more information aboutanomalies flagged in the Kepler pixel and flux time series.
KIC 3542116
Rap
papo
rt et
al.2
017,
arx
iv17
08.0
6069
A Search for Earth-size Planets
• Commissioning tools require special attention and data sets• Effort for commissioning tools may be as great as that for major science
pipeline modules• Don’t leave commissioning tool development to the last
Commissioning, Commissioning, Commissioning!
Kepler Commissioning TimelineVer. Flight17May 27, 2009B. Bensler
Calendar
Mission Events
Data Collection
SC Subsystems
GS Activities
Launch relative, L+
Day of MonthDay of WeekDay of Year
Month
Commissioning PhaseLaunch Phase Science Phase
State
ADCS
LGA Uplink Rate
Telecom
LGA Downlink Rate
HGA Downlink Rate
2 k125
7.8125
16 k2 k
100
3.465 M
2.475 M
Post-Launch Assessment
Review (PLAR)
Momentum Desats (C-010)
Acquire SunCoarse Point
Fine Point
Parameter File Updates
X-Band
Ka-Band
SSR Data Playback
Attitude
+Y to Sun
Science
Encircled Energy (C-031)
Uplink PRF Target Defs (C-038 / KACR-511)
Perform Preliminary CDPP (C-043)
Full Frame Images (FFI)
Long Cadence (LC) data sets(6) Gain and Linearity LCs (C-040)
Schedule Reserve
Data processingUplink processing (tables)
Data transfer (on ground, for major events)
Photometer Data
Characterize PRF (C-039)
(rates are expressed in bps)
Science Phase Readiness Review
xfr
(5) Scattered Light Sci FFI (C-035)
Uplink Target Defs (science) and Compression Tables (C-045 / KACR-550)
(121)EncircledEnergySCs(C-031)
2.887 M
4.331 M
2.165 M
Science Activity – Goal only
Ka-Band DownlinkDust Cover Ejection
(UTC)May 2009
Ka-BandDownlinks
Ka
xfr
Prime shift
xfr
ul
Configure for & Start Science (C-046)
ul
Science Ops
Ka Ka
Critical Path Activities in red font
(0.5 pixel offset)
May be used to make up actiivities
(121) Characterize PRF LCs (243 total), 15 min each (C-039)
Reference Pixels
Fine Guidance Sensor (FGS) FFI
(10) CDPP Reference Pixels (C-043, C-042)
11 x 11 pixel grid
Reverse Clocked (RC) data sets
Short Cadence (SC) data sets
Uplink PRF/CDPP TDTRef Pixels (C-038 / KACR-511)
(3) Ghosting Sci FFI (C-036)
(20 pixel offset)
AvionicsNVM Data Playbacks
Downlink Logs (C-048)
Encircled Energy Data Collection
Move FMs (KACR-423)
KaKa Ka
Focal Plane Geometry (C-032)
Scattered Light (C-035)
Ghosting (C-036)
(0.1 pixel offset)(0.1 pixel offset)
C-031C-032 C-039 C-043, C-045
FPG data processing
PRF data processing
Uplink science tables (ID = 013)
CDPP data processing
(0.5 pixel offset)
Characterize Gain and Linearity (C-040)
Project Reviews
20Mo110
17Fr
107
18Sa108
19Su109
13Mo103
14Tu104
15We105
16Th106
+37 +38 +39 +40 +41 +42 +43 +44 +45
27Mo117
24Fr
114
25Sa115
26Su116
21Tu111
22We112
23Th113
+46 +47 +48 +49 +50 +51 +52
1Fr
121
28Tu118
29We119
30Th120
+53 +54 +55 +56
June 09
Onboard Tables Set 1Set 2
5 of DiamondsPRF Targets (v-013) SciTDTs (v-020)
Telemetry ModeMode 6
Mode 3Mode 4
Mode 7
Mode 2Mode 1Mode 0
2 2 2 2 2 2 2 2
6 6 6 6 6 6 6 6
DSN CoverageGoldstone Viewperiod
Canberra Viewperiod
Madrid Viewperiod
>20 degElevation
Collect Ref. Pixels
= Active Table
Background Sequence
Ka
Star Alignment Determination
2
6
(3) RCs (C-043)
Ka
C-040
Gain and Linearity data goodness
Characterize PRF Ref Pixels(every other LC)
Thermal Stabilization
(3) RCs (C-043)(3) RCs(C-039) (3) RCs
(C-039)
Set 3
(3) RCs (C-046)
C-035, C-036
3.465 M2.887 M2.475 M2.165, 1.925, 1.732 M
4.331 M4.331 M 4.331 M 4.331 M 4.331 M 4.331 M
ul
PRF TDT uplink
CDPP TDT & Compression Table uplink
Science TDT uplink
EE OADR
PRFOADR
2Sa122
+57
6We126
3Su123
4Mo124
5Tu125
+58 +59 +60 +61
7Th127
X-Band
(1) FPGSci FFI(C-032)
(121)EncircledEnergySCs(C-031)
xfr xfr
Early SciOps Period(L+67 – L+90)
2
(1) Sci FFI(KACR-451)
Pre-Focus FFIs (KACR-451)Drop to Coarse Point (KACR-458)
Star Tracker Directed Searches (KACR-457)Return to Fine Point (KACR-460)
Data @ 4 corners (KACR-470)
(5) FGS FFI(KACR-470)
Multi-SCs(KACR-470)
Load Jitter TDTs (KACR-466)Collect Jitter LCs (KACR-465)
Ka
Multi-LCs(KACR-465)
Load FM blocks (KACR-474)
Collect EE data (KACR-423)
Collect EE FFI (KACR-476)
Optimize Focus (KACR-423)
Star Tracker Directed Searches (KACR-483)
Ka Ka
(1) Sci FFI(KACR-451)
(121)EncircledEnergySCs(C-031)(contingency)
Ka-Band Checkout (C-008 - Canberra)
4.331 M4.331 M
Ka
9Sa129
8Fr
128+62 +63 +64
13We133
10Su130
11Mo131
12Tu132
+65 +66 +67 +68
14Th134
+69
18Mo138
15Fr
135
16Sa136
17Su137
+70 +71 +72 +73
19Tu139
Random FFIs (KACR-469)
(5) Sci FFI(KACR-469)
Coarse Point (KACR-491)Load FM Blocks - Tilt (KACR-493)
Swap SSR Eng Partitions (KACR-489)
Move FMs - Tilt (KACR-493)Collect, playback 10 FGS FFIs (C-030a)
Ka
Update ST Alignment (C-030b)Collect, playback 10 FGS FFIs (C-030a)
ST Directed Search mods (KACR-502)
Ka
(20) FGS FFI (C-030a)
Update Science Attitude (KACR-504)
Return to Fine Point (KACR-504)Collect EE SCs and FFI (C-031)
(1) Sci FFI(C-031)
Ka
Reverse Clock5 of DiamondsEncircled Energy
6
2
4.331 M 4.331 M4.331 M
20We140+74 +75
21Th141
23Sa143
22Fr
142+76 +77 +78
27We148
24Su144
25Mo145
26Tu147
+79 +80 +81 +82
28Th149
+83
1Mo153
29Fr
150
30Sa151
31Su152
+84 +85 +86 +87
2Tu154
3We155+88 +89
4Th156
+90
5Fr
157
6Sa158+91
6
2
X-Band
CDPPTargets (v-014)
Uplink CDPP Target Defs (C-038 / KACR-518)Uplink Science tables (ID = 014)
Uplink Compression tables (ID = 200)
Disable XTWTA On Too Long Fault (KACR-548)Permanently Disable XTWTA On Too Long Fault (KACR-549)
Update LV commands in fault responses (KACR-534)
6
2
0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13 +14 +15 +16 +23 +24 +25 +26 +27 +28 +29 +30 +31 +32 +33
2 k
100 4.331 M
Launch Readiness Review (LRR)
COH Doppler
Photometer Initialization (C-014)
(8) Sci FFI
(C-018)
Enable sun avoidance (C-012)
Calibrate FGS & STA (C-030)
Eject dust cover (C-028)
Dust Cover Ejection Readiness Review
3.465 M2.887 M2.475 M2.165, 1.925, 1.732 M
6Mo96
7Tu97
8We98
3Fr93
4Sa94
5Su95
16Mo75
17Tu76
18We77
19Th78
21Sa80
22Su81
23Mo82
30Mo89
31Tu90
1We91
2Th92
6Fr65
7Sa66
8Su67
9Mo68
10Tu69
11We70
12Th71
13Fr72
14Sa73
15Su74
-1
Launch (~03:50 UTC for 93 deg azimuth)Injection and separationAuto-Init and de-tumble
Initialize SSR (C-007)
Rolling about sunline at 0.05 deg/sec
7.8125Ratefor NVM
playback
Stop Rotation
(temp)
FGS Mosaic
Ka
xf
Maneuver to science attitude (C-029)
20Fr79
2000
125
(3) Photometer checkout Sci FFIs (C-014)
(6) Photometer checkout FGS FFIs (C-014)
(36) Photometer checkout LCs (C-014)
First Contact (C-001)
Update parameters (C-003)Transition to Standby Mode (C-004)
S/C checkout, cmd check, enable COH Doppler, playback NVM
CSS Calibration (C-005)Point HGA to Earth (C-006)
X-band checkouts (C-002)
HGA Calibration (C-051)
Ka Ka
Measure Cosmic Rays (C-018)
Measure CCD Dark/Bias (C-019)
Measure FOV / First Light FFI (C-033)
(2) First LightSci FFI (C-033)
(10) Cal FGS & STA FGS FFIs (C-030)
Fine Point
Ka
C-030 C-030
9Th99
10Fr
100
11Sa101
12Su102
+34 +35 +36
Ka
222 2 2 21
3
7
3
0
6 6 6 6 6
xfrxf
C-033
!! LCP used for all Ka-band downlinks Switch to RCP
6
xfrxfr
C-003 C-005
1
7
C-030
(9) Sci FFIs
(C-019)
(9) RCs (C-019)
#1
#2 (also SDDL)
Ka
(9) Sci FFIs
(C-019)
(9) RCs (C-019)
Ka
#3
(9) Sci FFIs
(C-019)
(9) RCs (C-019)
C-019 1st set 48 hrs
xfr
xfr
Ka-band Checkout (C-008) - Madrid
Ka
0.275 M
#1
C-018
C-014
xfr
Ka-band Checkout (C-008) - Goldstone
(10) LCs (C-019)
(10) LCs (C-019)
4.331 M 4.331 M 4.331 M
6 6
22
3.465 M2.887 M2.475 M2.165, 1.925, 1.732 M
4.331 M 4.331 M2.887 M
C-019 4th set 84 hrs
Monitor 2D Black post DCE (C-053)
xfrC-053
OADR1 OADR2
CMC/SMD/Astro
16 k
CSS CalibrationAttitudes
Dark (#4)
(5) Ref Pixels
(C-019)
Maneuver to science attitude (C-029)
6
2
(2) LCs (C-019)
(1) Ref Pixels(C-019)
(2) FGS FFIs (Att #4)
(9) Sci FFIs
(C-019)
(9) RCs (C-019)
Ka
Dark (#4)
(9) Sci FFIs
(C-019)
(9) RCs (C-019)
(8) Sci FFI
(C-018)
(9) RCs (C-019)
(9) Sci FFIs
(C-019)
DCE CERR
NASA HQ PMC
4.331 M
C-019 2nd set 72 hrs
xfrC-019 3rd set 84 hrs C-019 5th/6th set
xfrC-018
xfr
Safe Mode Entry (KAR-419)
0
21
3
6
Return to Standby Mode (KAR-419)Power ON Photometer (KAR-419)
Photometer Aliveness Data (KAR-419)
ICV Load and Maneuver to SDDL (KAR-419)
KaKa
4.331 M
Patch Star Tracker (KACR-401)
Collect Science FFIs (KACR-385)
(2) FGS FFIs (KAR-419)
(1) Sci FFI(KAR-419)
(9) LCs (KAR-419)
(3) Sci FFIs(KACR-385)
xfrKACR-385
Patch Star Tracker (KACR-401)
ADCS Parameter updates (KACR-390, 394, 395)
Return to Standby Mode Attitude (KACR-402)
(31) LCs (KACR-407)
(10) Cal FGS & STA FGS FFIs (C-030)
(3) RCs (C-053)
(3) LCs (C-053)
(3) RCs (C-053)
(3) LCs (C-053)
(3) LCs (C-053)
(3) RCs (C-053)
6
2
April 2009March 2009
+17 +18 +19 +20 +21 +22
28Sa87
29Su88
24Tu83
25We84
26Th85
27Fr86
OADR3
A Search for Earth-size PlanetsPixel Response Function Characterization
synthetic image
480 500 520 540
550
560
570
580
590
600
610
flight image
480 500 520 540
550
560
570
580
590
600
610
Figure 2. A comparison of a synthetic image (left) with the actual flight image of the same region of the sky, near the edgeof the Kepler where the PRF is large. For purposes of comparison with the sky image the synthetic image is computedfor a short time interval for the time of the flight image so the PRFs are not smeared.
where νread is the value for this channel from the read noise and νquant is the quantization noise. Quantizationnoise is given by
νquant =
√
nC
12
( w
2nb−1
)2
(2)
where nC is the number of cadences in a co-added observation (30 for LC), w is the well depth and nb is thenumber of bits in the analog-to-digital converter (= 14). This noise formula (1) includes Poisson shot noise ofboth the target and background pixel values.
Given a collection of pixels, the SNR of the collection is given by the square root of the sum of squares of theSNR of the component pixels. Optimal pixel selection begins by including the pixel with the highest SNR. Thenext pixel to be added is the pixel that results in the greatest increase in SNR of the collection. Initially thecollection SNR will increase as pixels are added. After the bright pixels in the target have been added, dim pixelsdominated by noise cause the collection SNR to decrease. The pixel collection with the highest SNR defines theoptimal aperture. Fig. 3 shows an example of the dependence of SNR on pixel inclusion.
The background and target images allow us to estimate crowding by estimating the fraction of flux in theoptimal aperture that is due to the target star. The resulting crowding metric is useful for estimating the dilutionof flux from the target in the optimal aperture, which has an impact on the detectability of transits.7 The samemeasure on a 21×21 pixel square provides a sky crowding metric which is useful for identifying uncrowded stars.The fraction of each target’s flux that falls in its optimal aperture is also computed.
2.3 Background pixel selection
Background targets are selected to create a 2D polynomial representation of the background on each channel.This background polynomial is subtracted from the pixel values prior to aperture photometry.13 To lead togood polynomials the background targets should be homogeneously and uniformly distributed on the outputchannel. Because the accuracy of background polynomials increases with an increasing density of data pointsbut diminishes near the edge of the polynomial domain, more background targets are placed at the edges of theoutput channel. To achieve this distribution the intersections of an irregular Cartesian mesh are used for theinitial guess at background target positions. Because this mesh is the product of two linear meshes we cannotget a homogeneous distribution which provides exactly 1125 targets, but a choice of 31 × 36 = 1116 mesh linesis close.
synthetic image
480 500 520 540
550
560
570
580
590
600
610
flight image
480 500 520 540
550
560
570
580
590
600
610
Figure 2. A comparison of a synthetic image (left) with the actual flight image of the same region of the sky, near the edgeof the Kepler where the PRF is large. For purposes of comparison with the sky image the synthetic image is computedfor a short time interval for the time of the flight image so the PRFs are not smeared.
where νread is the value for this channel from the read noise and νquant is the quantization noise. Quantizationnoise is given by
νquant =
√
nC
12
( w
2nb−1
)2
(2)
where nC is the number of cadences in a co-added observation (30 for LC), w is the well depth and nb is thenumber of bits in the analog-to-digital converter (= 14). This noise formula (1) includes Poisson shot noise ofboth the target and background pixel values.
Given a collection of pixels, the SNR of the collection is given by the square root of the sum of squares of theSNR of the component pixels. Optimal pixel selection begins by including the pixel with the highest SNR. Thenext pixel to be added is the pixel that results in the greatest increase in SNR of the collection. Initially thecollection SNR will increase as pixels are added. After the bright pixels in the target have been added, dim pixelsdominated by noise cause the collection SNR to decrease. The pixel collection with the highest SNR defines theoptimal aperture. Fig. 3 shows an example of the dependence of SNR on pixel inclusion.
The background and target images allow us to estimate crowding by estimating the fraction of flux in theoptimal aperture that is due to the target star. The resulting crowding metric is useful for estimating the dilutionof flux from the target in the optimal aperture, which has an impact on the detectability of transits.7 The samemeasure on a 21×21 pixel square provides a sky crowding metric which is useful for identifying uncrowded stars.The fraction of each target’s flux that falls in its optimal aperture is also computed.
2.3 Background pixel selection
Background targets are selected to create a 2D polynomial representation of the background on each channel.This background polynomial is subtracted from the pixel values prior to aperture photometry.13 To lead togood polynomials the background targets should be homogeneously and uniformly distributed on the outputchannel. Because the accuracy of background polynomials increases with an increasing density of data pointsbut diminishes near the edge of the polynomial domain, more background targets are placed at the edges of theoutput channel. To achieve this distribution the intersections of an irregular Cartesian mesh are used for theinitial guess at background target positions. Because this mesh is the product of two linear meshes we cannotget a homogeneous distribution which provides exactly 1125 targets, but a choice of 31 × 36 = 1116 mesh linesis close.
Kepler PRF
TESS PRF
A Search for Earth-size Planets
Some Fast Code Some Slow Code
Keeping Up with the Data
A Search for Earth-size Planets
Step 1: Parallelize all code
Step 2: Make slow code fast(er)
Improving the Throughput
Some fast code; Some slow code
A Search for Earth-size Planets
64 hosts, 712 CPUs, 3.7 TB of RAM,148 TB of raw disk storage
Hardware Architecture: Kepler Science Operations Center
A Search for Earth-size Planets
5.34 Pflop/s peak cluster211,872 cores724 TB of memory15 PB of storage
Hardware Architecture: NAS Pleiades Supercomputer
A Search for Earth-size Planets
Characterizing completeness and reliability of software/people pipelines is extremely resource intensive Kepler shipped the final light curve products in April 2015We’ve spent the remainder of the time until present adding artificial transits, BEBs, scrambling the data temporally, inverting the light curves etc., etc.
hearth: Mapping Completeness and Reliability
Mapping completeness and reliability and characterizing the candidate vetting process is difficult
Recommendation: Pursue machine learning for conducting or modeling the candidate vetting process
A Search for Earth-size Planets
CALPixel level Calibration
PAPhotometric
Analysis
TPSTransiting Planet
Search
PDCPre-search Data
Conditioning
RawData
CorrectedLight Curves
Calibrated Pixels
Raw Light Curves & Centroids
DVData Validation
Diagnostic Metrics & Reports
Threshold Crossing Events(TCEs)
TCERTThreshold Crossing Event
Review Team+ Robovetting
Artificial Transit & BEB Injection Machine
Artificial Transits and Eclipses
Planet, or dud?Auto-Vetting
Applying machine learning to candidate
evaluation
>1,000,000 Lines of Code; 26 different Modules
Kepler’s Science Pipeline
A Search for Earth-size Planets
• ~13X pixel data rate over Kepler• Leveraged heritage from Kepler pipeline• Significantly lower cost (~46 FTEs over project lifetime)• Significant speed improvements:
• Colocated servers and storage with NAS Pleiades supercomputer• Moved pixel-level calibrations to C++• Sped up Presearch Data Conditioning by 10X• Originally projected 20+ days to process one sector• Complete pipeline requires ~5 days to process one sector
Developing the TESS Pipeline
SPOC
Science Pipeline & Other Infrastructure
PhotometricAnalysis
(PA)
Transiting Planet Search(TPS)
Raw Flux
DataValidation
(DV)
Threshold-crossing Events (TCEs)
Photometer Management
POC
OriginalPixel Data
POC
PlanetCandidates
and Reports
Photometer Performance Metrics
CDPP
Data Receipt
(DR)Archive
(AR)
Photometer Performance Assessment
(PPA)
Pipeline Infrastructure
(PI)
Presearch Data
Conditioning(PDC)
Raw Pixels
Datastore(ADB/Postgres)
Corrected Flux
CAL Metrics
TESS Input
Catalog
Target Lists
Calibration(CAL)
PA Metrics
Models (MOD)
Synthetic Data
OptimalApertures
Centroids
End-to-EndModel(Lilith)
Compression(COMP)
CompressionTables
Models
Compute Optimal Aperture
(COA)
CompressionTables
Exports
Target Lists
Synthetic Data
CalibratedPixels
Models
CompressionTables
TSO
MAST
ArchiveProducts
Models
Corrected Flux
Pipeline GUI
A Search for Earth-size Planets
The TESS Project is distributed geographically with the Science Pipeline separated by a continent from the Science Office’Resolving data issues requires good communication between the Payload Operations Center, the Science Processing Operations Center and the Science Office
Communication is KeyTESS Science Operations Center
SPOCScience Processing Operations Center (NASA Ames Research Center)
Processed Science Data
DV Products
Compression Tables
PPA Reports
Manifest
Acknowledgment
POCPayload Operations Center (MIT)
Science Data
Target Information
S/C Engineering Data
Focal Plane Models
Manifest
Acknowledgment
MASTMikulski Archive for Space Telescopes(STScI)
Acknowledgment
TSOTESS Science Office(MIT/Harvard)
TESS Input Catalog
TSO Data Products
Limb Darkening Model
Candidate Target List
Acknowledgement
NAIFNavigation and Ancillary Information Facility (JPL)
NAIF SPICE Kernels
FDFFlight Dynamics Facility(GSFC)
S/C Ephemerides
DSNDeep Space Network(JPL)
Ka-Band SFDU
TESSTransiting Exoplanet Survey Satellite (Space)
Instrument Data
Ka-B
and
SFDU
S/C
Ephe
m.
TSO Data Products
NAIF
SPI
CEKe
rnel
s
S/C Ephemerides
Manifest/Acknowledgment
Science Data
Target Info
Focal Plane Mod.
NAIF SPICE kernel
TIC, Limb-Dark.Models
Science Data/DV Products
Compression TablesPPA Reports
Manifest/Acknowledgment
Eng. DataFocal Plane Mod.Ephemerides
Sci. Data Comp.Tab.
TIC/ Limb.Dark.
Ack
Manifest
Science Data/DV Products
Cand.Targ. Lists
TICLimb-Dark. Models
Ack
Instr. Data
Ka-Band SFDU
Targ
et T
able
sCo
mp.
Tab
les
Target/Comp. Tables
Figure 2. The TESS ground segment is composed of the Science Processing Operations Center at NASA Ames ResearchCenter, the Payload Operations Center at MIT, the Mikulski Archive for Space Telescopes at STScI, the TESS ScienceO�ce in Cambridge MA, the Navigation and Ancillary Information Facility at JPL, the Flight Dynamics Facility atGSFC, and the Deep Space Network. This system manages the TESS science instrument to collect the science data andall the information necessary to carry out TESS’s main objective: searching for Earth’s nearest cousins. The diagramillustrates the flow of the information and data across the system.
The POC is the hub of the TESS ground segment, distributing information and data products throughoutthe ground segment as needed to support science operations. It passes the original science and engineeringdata to the permanent archive at the Mikulski Archive for Space Telescopes (MAST) located at theSpace Telescope Science Insitute (STScI) in Maryland, along with the archival science data products and allthe information necessary to process and analyze the science data, including ephemerides, and limb-darkeningmodels. The POC manages the science instrument and develops focal plane models critical for processing thescience data, including items such as the flat field, the pixel response function (PRF) and focal plane geometry(FPG). The POC accepts candidate target lists from the TESS Science O�ce (TSO), and formulates theflight target lists that determine which pixels are stored and downlinked from the spacecraft at the nominal2-minute science cadence.
The POC decompresses and delivers the original science and engineering data to the Science ProcessingOperations Center (SPOC) at NASA Ames Research Center in Mo↵ett Field CA, formatted into Googleprotocol bu↵ers, along with the focal plane models, target lists, and the target star information. The SPOCprocesses the original science data to produce calibrated pixels, light curves and centroids for all target stars,and associated uncertainties. The SPOC also calibrates the FFIs and provides world coordinate system (WCS)information for each frame. The SPOC develops compression tables based on commissioning data and sendsthese to the POC for formatting and uplinking to the spacecraft to achieve a compression performance meetingor exceeding 12 bits per science pixel. The SPOC searches the light curves for transiting planets and constructsa suite of diagnostic tests to make or break confidence in the planetary nature of each transit-like signature.The pixel, light curve, and transit search data products are sent to the POC, which transmits them to the
A Search for Earth-size Planets
New ideas for improving photometry/astrometry will emerge, both within the team and without
– “Halo” photometry on K2 data on Pleiades (White et al. 2017, MNRAS 471)
– “Everest” K2 photometry (Luger et al. arXiv:1702.05488)– Machine learning/Deep learning neural networks
Preserving ability to re-process the pixel data with better algorithms and tuned parameters is a really good idea
Take advantage of the compressibility of your data• Kepler achieved compression rates of 4.5 bits per pixel• TESS should achieve compression rates of ~3 bits per pixel for 2 minute
data and ~4 bits per pixel for 30 min FFIs
New Ideas for Every Step Will Emerge
A Search for Earth-size PlanetsSummary
• Science pipelines require significant planning and effort
• Previous pipelines can be leveraged to reduce development time (but this does not reduce time required for V&V testing)
• Plan to rewrite the majority of the science code in light of unexpected in-flight characteristics/behavior/hardware changes
• High fidelity simulations are indispensable
• Determining hearth is computationally intensive and huge effort
• Give adequate attention to developing commissioning scenarios and associated tools
• Take advantage of data compression to increase the amount of pixel data downlinked from PLATO
A Search for Earth-size PlanetsSOC Cluster Architecture
6 Clusters:4 Operations Clusters:
Flight Ops, Quarterly, Monthly& Archive)
2 Test Clusters:LAB & TEST
Science Processing PipelinesLong Cadence Photometry Pipeline
Transit Search Pipeline
TPS DV
CAL PA PDC PAD
FFI Pipeline
CAL PA
PPA
DYN TAD
Short Cadence Pipeline
CAL PA PDC
PMD PAG