1 of 11

Babamul & BOOM

Next Generation alert brokering

in the ZTF + Rubin era

Mansi M. Kasliwal (Caltech)

2 of 11

Team

Theophile Du Laz

Pete Bachant

Thomas Culino

Matthew Graham

Mansi Kasliwal

Michael Coughlin

Jake Simones

Matthew Jenson

Alexandra Junell

Argyro Sasli

Felipe Nunes

Maojie Xu

Benny Border

Antoine Le Calloch

Josh Bloom

Stefan Van Der Walt

+ Northwestern (Adam Miller)

+ University of Montclair (Shaon Ghosh)

3 of 11

BOOM

Multi-survey: process the ZTF + Rubin alert streams, with continuous cross-match to use their combined light curves at filtering time.
Data enrichment: Value added using ML scores using new multimodal models (Apple CiDeR), and xmatches with multiple static catalogs.
High performance: built in Rust for speed, on top of a non-relational database w/ native spatial querying support. Processing divided in multiple tiers/types of “workers”, to scale operations independently and horizontally.
User-defined: Filters are 100% customizable by users through the SkyPortal marshal, output sent their for manual and automated vetting+triggering.

4 of 11

ZTF

Kafka

Consumers

Database

ML Worker

Filter Worker

Kafka Topics

Alert Worker

Memory

Storage

LSST

+ …

HTTP Rest API

Marshals (SkyPortal/TOM)

BOOM’s

Architecture

archival + semi real-time searches

real-time + automated operations

Valkey over REDIS

NoSQL MongoDB

Rust (not Python)

5 of 11

babamul

Built on top of BOOM: Built on BOOM’s architecture and alert enrichment, it is an LSST filtering system.
Public filters: serving a set of static filters (based on transient/variable properties) based on 7+ years of ZTF operations
Low-latency: Filter results served over kafka, along with tools for the clients to filter further (e.g. local ML, custom catalog xmatch, …).
Multi-survey information: will include ZTF xmatch information, will include many catalog xmatch information, ML from AppleCiDEr

6 of 11

ZTF

Kafka

Consumers

Database

ML Worker

Filter Worker

Kafka Topics

Alert Worker

Memory

Storage

LSST

+ …

Public HTTP API

Public Clients

babamul’s

Architecture

7 of 11

Automated

Discoveries

Transient follow-up & classification
MMA event ToOs for GW/neutrinos
Enabled by BOOM + SkyPortal
All of the above, 100% automatic

8 of 11

Example: kilonova

9 of 11

BOOM

AppleCiDEr

Multimodal ML framework: photometry, image cutouts, metadata, spectra

Photometry: [CLS]-Transformer 87.8% accuracy (59.2% using the Informer)

Images+Meta: AstroMiNN slightly better performance compared to BTSbot

Spectra: SpectraNeXt-2D 87.6% accuracy (83.7% using the GalSpecNet)

AppleCiDEr: Averages the results of the above networks. Almost ready for production!

Interested? Please read Junell et al. 2025 (just submitted!)