AI system
- Model
-
BaselineCAE, a convolutional autoencoder, together with convolutional networks that classify glitches and signals. [1] [2]
- Release
-
The authors publish trained weights and a configuration dated 14 September 2026, without a separate model release number. [1] [2]
Editor’s note. The dated configuration comes from the authors’ repository changelog (reference 2). Their published checkpoint and code snapshot identify available files but do not establish the exact code commit used for every reported run.
- Role
-
Convolutional autoencoder for anomaly detection in the MADGRAV-2D cascade.
- Configuration
-
MADGRAV-2D combined anomaly detection with classifiers for glitches and low- and high-frequency signals. It analyzed observations from the Hanford and Livingston detectors during LIGO’s O3 and O4 runs. [1] [2]
- Inputs and outputs
-
Not assessed
- Autonomy
-
Not assessed
- Workflow
-
Sequential CNNs for anomaly detection, glitch classification, coherence testing and signal ranking on one-second Q-transform windows of Hanford and Livingston data. 442.9 days of coincident livetime across O3a, O3b, O4a and O4b.
- Human oversight
-
Not assessed
- Relation to AI
-
The convolutional networks performed the whole detection chain, from anomaly detection to ranking, and the authors are the developers. Every detection matches a published catalog event. The authors state that no candidate without a catalog counterpart survives, so the search added no new discovery in this analysis.
- AI contribution
-
The convolutional networks performed the whole detection chain, from anomaly detection to ranking, and the authors are the developers. Every detection matches a published catalog event. The authors state that no candidate without a catalog counterpart survives, so the search added no new discovery in this analysis.
Benefits
- Outcome identifiers
-
Not assessed
- Benefit
-
47 detections with a false alarm rate below one per year and p_astro above 0.9
- Category
-
Improved Task Accuracy
- Status
-
Not assessed
- Beneficiary groups
-
Organizations. Society.
- Beneficiaries
-
Not assessed
- Reach
-
Not assessed
- Distribution and access
-
Not assessed
- Duration
-
Not assessed
- Expected?
-
Yes
- Costs and harms
-
Not assessed
Measurement
- Studies
-
Not assessed
- Study design
-
Not assessed
- Measurement
-
Not assessed
- Setting and sample
-
Not assessed
- Measures
-
47 detections with a false alarm rate below one per year and p_astro above 0.9
- Comparison
-
the coherent WaveBurst (cWB) search and the published catalogs on the same LIGO data
- Results
-
Result: 47 detections, 17% of the catalog population and 39% of the cWB count.
- Additional value from AI
-
Not assessed
- Limitations
-
Not assessed
- Denominators
-
Not assessed
- Materials for replication
-
Not assessed
Money
- Kind of value
-
Not assessed
- Amount
-
Not assessed
- Currency and prices
-
Not assessed
- Period and unit
-
Not assessed
- Realization
-
Not assessed
- Recipients and payers
-
Not assessed
- Gross or net
-
Not assessed
- Included costs
-
Not assessed
- Comparison
-
Not assessed
- Method
-
Not assessed
- Return and payback
-
Not assessed
People and places
- Participants
-
Developer: Marietta Blau Institute for Particle Physics, Austrian Academy of Sciences, with the University of Urbino and INFN Florence. Operator: The same authors, on public LIGO O3 and O4 data.
- Jurisdictions
-
The research team had Austrian and Italian affiliations and used computing facilities in
Austria. LIGO instruments in the 
United States supplied the observations. [1] - Ownership and control
-
Not assessed
- Dependencies
-
Not assessed
- Cross-border connections
-
Not assessed
- Geopolitical implications
-
Not assessed
Follow-up
- Conditions for replication
-
High-mass, high-SNR sources: the recovered fraction rises from 8.1% below 30 solar masses to 53.3% above 100. Two-detector coherence between Hanford and Livingston.
- Replication attempts
-
Not assessed
- Later findings
-
The authors revised the preprint on 19 October 2026. No independent follow-up result is cited. [1]
- Availability
-
Not assessed
- Institutional decision
-
Not assessed
- Grounds
-
Not assessed
- Action and result
-
Not assessed
- Success criterion
-
Not assessed
Evidence
- Source numbers
-
AIBD-0283-source-0001–0002
- Citations
-
Numbered references below. [1]
- Language
-
Not assessed
- Source relationships
-
Developer [1]
- Original evidence
-
The developers’ preprint reports the detections. Their repository supplies the code, trained weights, and configuration. The paper’s revisions and code are supporting material from the same team. [1] [2]
- Supported findings
-
Not assessed
- Evidence links
-
Not assessed
- Evidence strength
-
Not assessed
- Missing evidence
-
Not assessed
Grades
- Assessed benefits
-
Organizations. Society.
- Depth
-
Not assessed
- Reasons
-
Organizations: Recovery of 47 known events, 44 of them shared with coherent WaveBurst, demonstrates the method. No new event, catalog adoption or operational use is documented. Society: A complementary search channel is proposed; no scientific result beyond re-detection is claimed.
- Observed or intended
-
Not assessed
- Rubric
-
Not assessed
- Assessor and date
-
Not assessed
- Alternative assessments
-
Not assessed
Entry details
- Kind
-
Not assessed
- Related entries
-
Not assessed
- Field
-
Astrophysics / gravitational-wave astronomy. Academic research on public detector data.
- Purpose
-
MADGRAV was built to search LIGO data for high-mass compact binary coalescences by anomaly detection. Recovering cataloged events was that purpose.
- Dates
-
Source published: 30 September 2026. Updated: 25 October 2026. [1]
- Stage
-
Not assessed
- Version
-
SHA-256: 9a239a727d55…
References
- Other identifiers
-
Not assessed
- Category mappings
-
Not assessed
- Brief
-
Not assessed
- Revisions
-
15 revisions. Latest: 25 October 2026. [1]
Sources
-
Preprint Developer
Inguglia G, Haigh H, Dupletsa U, Longo A (2026). MADGRAV: a multilevel anomaly-detection pipeline for gravitational-wave searches applied to LIGO data. arXiv:2609.39583v1, 30 September 2026, abstract and sections 3–5. Revised as v2 on 19 October 2026. Authored by the pipeline's developers.
Version 2 author affiliations establish Austria and Italy, and acknowledgments identify Austrian computing facilities. Author affiliations and Acknowledgements, paragraph 1.
One primary preprint by the pipeline developers. Its abstract, HTML versions, and revision are the same work, not independent evidence.
Of the 47 detections, 44 are shared with the minimally modelled coherent WaveBurst search.
-
Other Developer
Original code, trained weights, and frozen configuration linked from section 2 of the paper. The repository identifies MADGRAV-2D and the final configuration that produced the reported detections. This supplies software provenance, not independent corroboration.
This is the final version of the pipeline