TBD_LLNL_proposers_day_briefing.pdf

1LLNL-PRES-2013156

Prepared by LLNL under Contract DE-AC52-07NA27344.

Astronomy and Astrophysics Analytics Group

Space Program

Dr. Nathan Golovich

4 November 2025

Track at Big Distances with Track-Before-Detect (TBD2) Proposer’s Day

Lawrence Livermore National Laboratory

Distribution Statement A: Approved for public release. Distribution unlimited.

2LLNL-PRES-2013156

• Brief introduction to LLNL

• Background on track before detect (TBD) methods

• Considerations for TBD in space

• Primer on Livermore Computing (LC)

Outline

3LLNL-PRES-2013156

• Department of Energy managed

FFRDC

• National Nuclear Security Administration (NNSA) laboratory

• 9500 employees, $3.25B budget

• Scientists and Engineers working on mission focused R&D

• World’s most powerful computer and laser

Livermore overview

4LLNL-PRES-2013156

• Leverages HPC capabilities

• Wide field surveys, big data problems

• Rubin/ZTF

• Track before detect at large scales

• Large scale modeling and simulation

• Space payload and mission development

• Several active and planned missions from LEO to lunar orbit, often COTS hardware and commercial spacecraft

• Leverages NIF and advanced manufacturing expertise of the lab

• Precision X-ray calorimeters

• Adaptive optics for ground and future space based missions

Livermore space and astronomy Work

5LLNL-PRES-2013156

• Natural extension to image coaddition for increasing SNR of faint objects

• For N images of an object the noise scales as 1/sqrt(N)

• Methods account for motion against registered frames

• Typically means sidereal tracking and astrometric frame registration to star catalogs

• First methods for slow moving outer solar system bodies

Track before detect methods 1

Non-linear residual motion for various solar system bodies Heinze et al. (2016)

6LLNL-PRES-2013156

• TBD enabled the discovery of the faintest and furthest objects in the outer solar system (20-40 km objects outside Neptune)

• Methods required additional non-linear motion parametrization and weeks of computing on powerful (circa 2004) workstations

Track before detect methods 2

Bernstein et al. (2004)

7LLNL-PRES-2013156

• Extensions to TBD methods have been implemented at

LLNL

• Extending to wide field, long time-domain surveys like ZTF and Rubin

• Requires full 6d motion model

Track before detect methods 3 8405 Asbolus (1995 GO) 𝑎𝑎 = 18 AU

See Golovich et al. (2025) and Geringer-Sameth et al. (2025)

8LLNL-PRES-2013156

• Method inherently friendly to parallelization and acceleration

• Options for preprocessing for streamlined pipelines, load sharing between compute arcitectures

• False positives are a real challenge

• Large number of trials sampling pixels randomly means rare events are likely to occur

• One in a million is not unlikely when you try ten million times!

• This may mean that SNR thresholds must be higher than we’re used to

• Empirical motion is derivable given the geometry of the observation and the trajectory of the hypothesized target

• Provably optimal signal-matched-filtering

• Sideral tracking is ok if single frame exposure times are short compared to motion of targets

• Minimal trailing losses, easier astrometry

• Challenge: need to get enough stars

• Difference images are your friend

Track before detect methods 4: nuts and bolts

9LLNL-PRES-2013156

• Imaging from Sun-Earth L1 has plusses and minuses

• + Target is fully illuminated

• + All of cislunar is compact in angular extent

• - Earth and Moon are fully illuminated

• - Only a single perspective, difficult to measure distance and radial velocity

• Imaging from a small satellite is not as stable as imaging from a ground telescope

• Jitter and drift can be large, forcing short snapshots for each frame

• Read noise dominates zodiacal lighte

• SWaP

• Power constraints limit compute

• What shortcuts can we make, at what cost to sensitivity?

Considerations for TBD in space

10LLNL-PRES-2013156

Tracking 2005 LW3 with GEOstare SV2

Raw data ~200 minute period

We detected the asteroid in 49 epochs over 1.5 days

Demonstration of small-satellite tracking / monitoring capability

U nf ilt er ed m ag ni tu de

U nf ilt er ed m ag ni tu de

Minutes since Wed Nov 23 10:06:43 UTC Minutes since Wed Nov 23 10:06:43 UTC

Significant brightness increase in 25 minutes

Periodogram / folded period

6u cubesat with 8.5cm LLNL monolith optic

11LLNL-PRES-2013156

• For TBD2, LLNL will simulate realistic space-based image stacks for performers

• Features:

• Use GalSim injected star/galaxy fields

• Random cislunar objects in various orbits (DRO, L1/2 halo orbits, etc)

• Realistic read-noise dominated pixel noise

• Realistic geometry and timing provided by LLNL

SSAPy orbit propagation tool

• Explore realistic noise profiles from Earth/Moon

• We will also explore providing space data from GEOStare SV2

LLNL Simulation of Image Stacks

Example of a simulated 3 Gpix simulated Rubin/LSST image with injected satellites

12LLNL-PRES-2013156

Livermore Computing

Main supercomputing facility: 48,000 square feet of unobstructed computer room floor, 30 megawatts of machine power capacity, and 7,200 tons of air conditioning capacity.

Several auxiliary facilities on campus

• Exaflop supercomputer El Capitan,

• Commodity clusters with GPUs

• Emerging architectures

• Cerebras wafer scale supercomputer

13LLNL-PRES-2013156

• Remote power management and monitoring capability

• Access control

• Physical: separate networks for performers

• Access control lists and user login restrictions enforced

• Support for dedicated native OS and drivers

• Root access privileges if needed

• Enrollment through Livermore Computing collaboration process

• Unclassified use

• Dedicated network populated with compute nodes and DARPA-selected devices

• Scheduled dedicated use for performers

• Repeatable experiments

• Customized measurements

• Support for Docker containers

• Locally hosted data sets provided by LLNL and JPL

LLNL testbed for measurement and validation

TBD2 mid term and final evaluation will involve running pipelines on simulated and real image stacks on the LLNL testbed. Results will be shared with LLNL and JPL for validation