Jenkins PLATO Lessons Learned 2017 · 2019. 8. 30. · Characterize PRF(C-039) (rates are expressed in bps) Science Phase Readiness Review xfr (5) Scattered Light Sci FFI (C-035)

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



• 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


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


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


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


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


• 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


(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)


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


Some Fast Code Some Slow Code

Keeping Up with the Data


Step 1: Parallelize all code

Step 2: Make slow code fast(er)

Improving the Throughput

Some fast code; Some slow code


64 hosts, 712 CPUs, 3.7 TB of RAM,148 TB of raw disk storage

Hardware Architecture: Kepler Science Operations Center


5.34 Pflop/s peak cluster211,872 cores724 TB of memory15 PB of storage

Hardware Architecture: NAS Pleiades Supercomputer


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


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


• ~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


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


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

Documents

Jenkins PLATO Lessons Learned 2017 · 2019. 8. 30. · Characterize PRF(C-039) (rates are expressed in bps) Science Phase Readiness Review xfr (5) Scattered Light Sci FFI (C-035)