Science Alert Generation in CTAO

• Science Alert Generation (SAG) is an automated pipeline in CTAO’s ACADA suite that rapidly identifies, analyzes, and communicates very-high-energy gamma-ray transients.
• It integrates data reconstruction, quality control, and scientific analysis subsystems to produce significance estimates, flux measurements, sky maps, and light curves.
• SAG achieves ultra-low latencies (5–20 seconds) through parallel processing, database-driven orchestration, and GTI-based filtering for robust and timely transient alerts.

Science Alert Generation (SAG) is the automated pipeline within the Cherenkov Telescope Array Observatory (CTAO) Array Control and Data Acquisition (ACADA) software suite for the real-time identification, scientific analysis, and communication of transient and variable very-high-energy gamma-ray phenomena. It performs the first real-time scientific analysis after data acquisition, processes data on multiple time scales from seconds to hours, and is designed to issue candidate science alerts for rapid follow-up by internal systems and the wider astronomical community (Bulgarelli et al., 2021, Panebianco et al., 23 Sep 2025). In current CTAO usage, SAG denotes an observatory-integrated alerting system composed of supervisory, reconstruction, data-quality, and science-analysis subsystems, with SAG-SCI providing the real-time, high-level analysis that turns reconstructed gamma-ray events into significance estimates, flux measurements, sky maps, light curves, blind-search results, and candidate alerts (Collaboration et al., 19 Sep 2025).

1. Observatory context and operational role

CTAO is intended to be the leading observatory for very-high-energy gamma-rays over the next decades, and its large effective area, field of view, sensitivity, and rapid slewing capability make it especially suited to the study of transient astrophysical phenomena (Panebianco et al., 23 Sep 2025, Bulgarelli et al., 2021). The CTA system is designed to detect and react to gamma-ray bursts, flaring active galactic nuclei, gamma-ray binaries, serendipitous transient events, and counterparts to gravitational-wave or neutrino triggers (Fioretti et al., 2019, Bulgarelli et al., 2021).

Within ACADA, SAG is the subsystem dedicated to automatic science alert generation. The broader ACADA system controls, supervises, and handles the data generated by the telescopes and auxiliary instruments, while the Transients Handler manages both science alerts produced by SAG and external alerts received from other instruments (Oya et al., 17 Jun 2026). The Short-Term Scheduler can then modify ongoing observations on sub-minute timescales in response to relevant alerts (Oya et al., 17 Jun 2026). This coupling of rapid response, real-time analysis, and fast telescope slewing is central to CTAO’s time-domain and multi-messenger observing model.

CTA studies describe an evolution in latency targets. A transient-sensitivity study referred to an on-line SAG pipeline with a required 30 second latency (Fioretti et al., 2019). Later CTAO architecture papers specify that SAG must issue candidate science alerts within 20 seconds from the data taking and that SAG-SCI itself is designed to issue alerts within 5 seconds of data availability (Bulgarelli et al., 2021, Panebianco et al., 23 Sep 2025). This suggests a progressive tightening of real-time performance requirements as the architecture matured.

2. Component structure inside ACADA

SAG is partitioned into four principal components, coordinated inside ACADA and instantiated per active sub-array where required (Bulgarelli et al., 2021, Collaboration et al., 19 Sep 2025).

Component	Stated function	Operational detail
SAG-SUP	Supervises the pipeline execution	Oversees dynamic operations and integration with other ACADA components
SAG-RECO	Handles Cherenkov data reconstruction	Performs fast, online event reconstruction from DL0 to DL3
SAG-DQ	Performs data quality checks	Monitors raw to reconstructed data and can generate DQ vetoes
SAG-SCI	Performs high-level, real-time scientific analysis	Monitors sources, detects variability or transients, and generates candidate science alerts

SAG-RECO performs the fast conversion of telescope data into reconstructed gamma-ray-like events. In CTA descriptions, this includes image integration, image cleaning, parameter extraction, directional and energy reconstruction, gamma/hadron discrimination, and event selection, with machine learning used for rapid event characterization and with configuration-matched instrument response functions selected for the active sub-array (Bulgarelli et al., 2021, Collaboration et al., 19 Sep 2025).

SAG-DQ provides the data-quality gate on which alert validity depends. It performs online monitoring from raw to reconstructed data, assesses instrument and data health, and can issue warnings, alarms, or vetoes on problematic intervals or events (Baroncelli et al., 2021, Collaboration et al., 19 Sep 2025). The rta-dq-lib library proposed for this role is available in Python for rapid prototyping and in C++ for production-grade performance; it is configured through XML and internally translated into a directed acyclic graph of DataSource, Executor, QualityCheck, and Output nodes (Baroncelli et al., 2021).

SAG-SUP is the orchestration layer. In the newer architecture it comprises a singleton SAG Supervisor, dynamically spawned SAG Pipeline Sub-array Supervisors, and the SAGDataObserver introduced in Rel2 (Collaboration et al., 19 Sep 2025). The supervisors manage lifecycle transitions, collect telescope status and environmental information, coordinate SAG-RECO, SAG-DQ, and SAG-SCI, and contribute to the computation of Good Time Intervals for trustworthy alert generation (Collaboration et al., 19 Sep 2025).

3. Data ingestion, supervision, and validity filtering

The SAG-SCI workflow is database-driven and modular. After observation, gamma-ray event lists produced by SAG-RECO are generated for each run. A daemon called DL3Merger monitors new data availability, reads event list files, and merges them into a central MySQL Data Source database, together with associated metadata such as observation parameters and instrument specifics (Panebianco et al., 23 Sep 2025). A separate Data Model Database maintains information on runs, instruments, target sources, analysis jobs, and analysis types, separating static from dynamic content, while Science Logic implements rules and triggers, using MySQL triggers, to schedule analyses as data arrive (Panebianco et al., 23 Sep 2025).

Parallel execution is handled by three Python daemons—DL3Merger, SubmitJob, and UpdateJobStatus—together with the SLURM Workload Manager (Panebianco et al., 23 Sep 2025). This arrangement allows multiple analyses to run simultaneously and supports prioritization of rapid-response analyses, especially during Target of Opportunity follow-ups (Panebianco et al., 23 Sep 2025). In the broader CTAO architecture, SLURM also supports sub-array-level parallelism and horizontal scaling (Bulgarelli et al., 2021, Oya et al., 17 Jun 2026).

A central architectural development is the explicit filtering of data according to telescope state, environmental conditions, and data-quality status. SAG-SUP can retrieve real-time telescope status and environmental conditions from telescope managers and the weather station through the ACADA Monitoring system, collect them in a database, and use them to filter out data from slewing phases or degraded conditions (Collaboration et al., 19 Sep 2025). Rather than waiting for all telescopes in a sub-array to reach stable tracking, SAG-SUP records individual transition times through startDataTaking(TelescopeID) and enables analysis to begin as soon as valid tracking data are available (Collaboration et al., 19 Sep 2025).

This filtering is formalized through Good Time Intervals (GTIs). Standard GTIs are based on telescope tracking status,

$$\text{GTI}_{\text{std}} = \bigcup_{\text{telescopes}} \left[ t_{\text{start,track}}, t_{\text{end,track}} \right],$$

while enhanced GTIs additionally intersect telescope-status validity with environmental constraints and DQ vetoes,

$$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$

Events outside valid GTIs are discarded or flagged by SAGDataObserver before they reach SAG-SCI (Collaboration et al., 19 Sep 2025). A plausible implication is that alert validity in CTAO is defined not only by statistical excess but also by formal interval-level admissibility.

4. Real-time scientific analyses in SAG-SCI

SAG-SCI is the subsystem responsible for continuous scientific monitoring of all observed sources and for generating candidate science alerts when flares, new sources, or significant deviations from baseline behaviour are detected (Panebianco et al., 23 Sep 2025). Its pipelines receive gamma-ray data from multiple reconstruction lines, merge them, store them in a database, and trigger several parallel scientific analyses on the latest data (Panebianco et al., 23 Sep 2025). These analyses are implemented through Python-wrapped science tools, primarily Gammapy and a custom wrapper called RTAPH, within a modular framework that supports prioritization and concurrent execution (Panebianco et al., 23 Sep 2025).

The analysis portfolio is broader than simple threshold crossing. SAG-SCI estimates target significance and flux, produces sky maps and light curves, and conducts blind searches for sources within the field of view (Panebianco et al., 23 Sep 2025). Light-curve computation uses a sliding-window approach, with time bins as short as 10 seconds and as long as 180 minutes, and results over multiple bins can be cumulatively combined by stacking to increase sensitivity to faint or persistent sources (Panebianco et al., 23 Sep 2025). Sky analyses include sky count maps, significance maps over the field of view, and $\theta^2$ plots comparing photon distributions around the source versus control regions (Panebianco et al., 23 Sep 2025).

Blind-search capability is a defining feature. Hotspots in the field of view that are not associated with catalogued sources and that cross the significance threshold can trigger alerts, after which a new dedicated targeted analysis of the spot is launched for dedicated flux and light-curve estimation (Panebianco et al., 23 Sep 2025). Source cross-matching with very-high-energy catalogues is used to distinguish known from unknown sources (Panebianco et al., 23 Sep 2025). This makes SAG-SCI simultaneously a source-monitoring system, a variability detector, and a serendipitous discovery engine.

5. Alert criteria, statistical methods, and uncertainty budgets

A candidate science alert is generated by SAG-SCI when the statistical significance and the detected flux of a source exceed predefined thresholds, when there is a significant deviation from known reference states such as catalog fluxes, or when a previously unknown or serendipitous source is detected with high significance (Panebianco et al., 23 Sep 2025). For known sources, both flaring and low states can satisfy the alert logic if the measured flux is notably higher or lower than the historical reference (Panebianco et al., 23 Sep 2025). For new or poorly localized transient sources, detection without a catalogue counterpart is sufficient for alerting (Panebianco et al., 23 Sep 2025). Alerts are not issued, or are flagged, if SAG-DQ detects anomalies in the input data (Panebianco et al., 23 Sep 2025).

SAG-SCI quantifies detection significance through On/Off region analysis with background estimation via aperture photometry or reflected regions, and through likelihood-ratio methods (Panebianco et al., 23 Sep 2025). In sky-map style analyses, significance is computed as

$\text{Significance} = \sqrt{\rm TS},$

with

$TS = 2 \times (\log L_1 - \log L_0),$

where $L_1$ is the likelihood with a source and $L_0$ is the likelihood under the background-only hypothesis (Panebianco et al., 23 Sep 2025). Earlier CTA sensitivity work examined the same two methodological families—aperture photometry and likelihood analysis—as the algorithms that define CTA’s short-exposure transient sensitivity (Fioretti et al., 2019). In that study, the On-Off method used a typical $\alpha=0.2$, required $N_{\rm s} \geq 10$, and imposed a background-systematics condition, while a typical $\mathrm{TS}=25$ corresponded to $$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$0 significance in likelihood analysis (Fioretti et al., 2019).

Flux estimation in SAG-SCI is based on maximum-likelihood fits to binned gamma-like events within a time window for a spectral model, usually a power law,

$$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$1

In real-time operations, the current approach assumes a fixed spectral index, usually $$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$2, and fits only the normalization $$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$3 (Panebianco et al., 23 Sep 2025). The dominant SAG-SCI-specific systematic arises from errors in this assumed spectral index: for a 20% under- or over-estimation in spectral index, the resulting systematic error on flux can reach approximately 30% (Panebianco et al., 23 Sep 2025). For 100-second timescales, the statistical errors on flux are typically about 35%, primarily because of limited photon statistics (Panebianco et al., 23 Sep 2025). These values define the current real-time uncertainty budget more directly than generic statements about alert sensitivity.

6. Validation results, system performance, and wider context

SAG-SCI has been validated on simulated gamma-ray data in three main scenarios: steady source monitoring, transient flaring source detection, and blind search for serendipitous transients (Panebianco et al., 23 Sep 2025). In the steady-source case, the system detected the source and reconstructed the average flux in agreement with the simulated value within the quoted statistical and systematic errors (Panebianco et al., 23 Sep 2025). In the flaring-source case, a source with a fast, 10x exponential flux flare was detected in the correct time bin—the fifth bin of a 100-second binned light curve—with significance greater than $$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$4 and flux about 10 times the quiescent value (Panebianco et al., 23 Sep 2025). In the blind-search case, the algorithm correctly identified an off-axis transient at the correct offset, mapped its position together with that of a steady source, and measured a significance of $$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$5 in a short exposure (Panebianco et al., 23 Sep 2025).

These simulation results are aligned with the stated low-latency design. SAG-SCI is designed to issue alerts within 5 seconds of data availability (Panebianco et al., 23 Sep 2025), while the full SAG system must issue candidate science alerts within 20 seconds from data taking and operate on trigger rates of tens of kHz (Bulgarelli et al., 2021, Collaboration et al., 19 Sep 2025). Earlier CTA transient studies had framed the corresponding requirement as 30 seconds from acquisition of the last relevant event (Fioretti et al., 2019). The architecture therefore couples a stringent timing budget with concurrency, database-mediated orchestration, and GTI-based validity control rather than with a single monolithic analysis.

The term “SAG” also has established meanings outside CTAO. In large-scale optimization it denotes Stochastic Average Gradient, and in computer vision it denotes Self-Assessed Generation for optical flow and stereo matching (Zhu et al., 5 Feb 2026, Ling et al., 2024). In high-energy astrophysics, however, CTAO’s Science Alert Generation pipeline belongs to a broader class of automated transient-alert systems. The AGILE Science Alert System, for example, used distributed alert pipelines, blind-search and multi-source likelihood analysis, and a false discovery rate threshold of $$\text{Enhanced GTI} = \text{Standard GTI} \cap \text{[Environment OK]} \cap \text{[DQ OK]}.$$6 for one of its gamma-ray transient searches (Trifoglio et al., 2013). This suggests that SAG should be understood as part of a longer lineage of rapid alert infrastructures, while remaining specific in its integration with ACADA, its use of sub-array supervision and GTIs, and its requirement to combine reconstruction, data-quality analysis, and science analysis in a single real-time observatory framework (Collaboration et al., 19 Sep 2025, Oya et al., 17 Jun 2026).

Current development directions are explicitly identified in CTAO materials. Future versions aim to reduce systematics by enabling real-time spectral-index fitting, background modeling, and improved catalogue cross-matching (Panebianco et al., 23 Sep 2025). Given the present architecture, a plausible implication is that subsequent gains in scientific performance will depend not only on faster reconstruction or larger compute budgets, but also on tighter coupling between statistical inference, data-quality veto logic, and observatory-level scheduling.