Localization and Confidence Region Estimation of Short GRBs with the COSI BGO Shield Using a HEALPix-Based Deep Learning Approach

N. Parmiggiani , A. Bulgarelli , G. Panebianco , E. Burns , E. Neights , V. Fioretti , I. Martinez-Castellanos , L. Castaldini

show 12 more authors

A. Ciabattoni A. Di Piano R. Falco S. Gallego G. Mustafa P. Patel A. Rizzo E. A. Wulf D. H. Hartmann C. A. Kierans J. A. Tomsick A. Zoglauer

Authors on Pith no claims yet

Pith reviewed 2026-05-10 10:04 UTC · model grok-4.3

classification 🌌 astro-ph.HE astro-ph.IM

keywords short gamma-ray burstsGRB localizationdeep learningHEALPixanticoincidence shieldconfidence regionsBGO detector

0 comments

The pith

A neural network outputs sky probability maps from BGO shield counts to localize short GRBs and derive their 90% confidence regions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a deep learning method that localizes short gamma-ray bursts detected by the anticoincidence subsystem shields of the COSI instrument. It trains a classifier to produce a probability distribution over the sky in the HEALPix pixelization, from which 90% confidence regions are extracted even when those regions consist of multiple disconnected patches. This matters for an automated ground pipeline that must deliver positions quickly after an onboard trigger. A sympathetic reader would see value in a technique that turns raw count patterns into usable sky constraints without requiring full directional reconstruction.

Core claim

A neural network classifier applied to ACS data predicts the GRB location as a probability distribution across a HEALPix sky map, allowing direct computation of 90% confidence regions that may be split into separate areas.

What carries the argument

The HEALPix-based neural network classifier that converts shield count rates into a sky probability distribution used to define confidence regions.

If this is right

The method can localize GRBs and report 90% regions even when the high-probability area is disconnected.
The probabilistic output supports automated pipelines that must share positions with the community after each trigger.
The approach is positioned for future direct comparison against classical chi-squared fitting and maximum-likelihood estimation on the same data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the network performs well on real data, the same probability-map output could be combined with localizations from other instruments to tighten multi-messenger alerts.
The HEALPix representation may allow straightforward integration of the shield-based constraints into broader sky-survey pipelines that already use equal-area pixelizations.
Similar classifier architectures could be tested on other wide-field detectors that record only coarse count rates rather than full imaging.

Load-bearing premise

A network trained on simulated or labeled ACS data will generalize to real short GRBs observed by the flight instrument without loss of localization accuracy or confidence coverage.

What would settle it

For a sample of short GRBs with independently measured true positions, count how often those positions fall inside the predicted 90% regions; systematic under-coverage below 90% would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.15119 by A. Bulgarelli, A. Ciabattoni, A. Di Piano, A. Rizzo, A. Zoglauer, C. A. Kierans, D. H. Hartmann, E. A. Wulf, E. Burns, E. Neights, G. Mustafa, G. Panebianco, I. Martinez-Castellanos, J. A. Tomsick, L. Castaldini, N. Parmiggiani, P. Patel, R. Falco, S. Gallego, V. Fioretti.

**Figure 1.** Figure 1: Left: COSI ACS Design. Right: BGO crystal scheme. 2. Methods We simulated two sGRB datasets using MEGAlib (Zoglauer et al. 2006) for training and testing purposes. The simulation procedure is described in detail in Parmiggiani et al. (2026). The testing dataset contains 49,152 sGRBs distributed following the HEALPix coordinate system (nside=64), while the training dataset contains 100,000 sGRBs randomly s… view at source ↗

**Figure 2.** Figure 2: Examples of 90% c.l. error regions (997 deg [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

read the original abstract

The Compton Spectrometer and Imager is a NASA satellite mission under development that will survey the entire sky in the 0.2-5 MeV range using a wide-field germanium detector array, surrounded on the sides and bottom by active shields (the Anticoincidence Subsystem, ACS). The ACS aims to suppress and monitor background events, as well as detect transient sources, such as Gamma-Ray Bursts (GRBs), through its onboard triggering algorithm. The data related to GRBs are sent to the ground and analyzed by an automated pipeline to localize the GRBs and share their positions with the community. In this work, we present a brief GRB localization method based on ACS data, utilizing deep learning (DL) techniques, which can estimate the 90\% confidence region, including cases where it is split into multiple areas. To address this, we developed a neural network classifier that predicts the GRB location as a probability distribution across the sky map following the HEALPix framework. The distribution can be used to compute the 90\% confidence regions. Future work will compare this DL-based localization approach with classical methods such as $\chi^2$ fitting and Maximum Likelihood Estimation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

read the letter

The paper sketches a HEALPix neural-net pipeline for turning COSI BGO shield counts into 90% confidence regions that can be disconnected, but supplies no metrics, training details, or tests on real or realistic data. That is the main thing to know: it is a methods outline aimed at the COSI ground pipeline for short GRBs, not a validated tool yet. The central technical move is training a classifier to output a full sky probability map on the HEALPix grid instead of a single best-fit point or ellipse. From the map the authors extract the smallest region containing 90% probability, which can split into separate patches when the shield response is ambiguous. This is a practical response to how anticoincidence data often produces multi-modal likelihood surfaces. The choice of HEALPix is straightforward and fits the all-sky, uniform-pixel requirement. The work is coherent on its own terms and shows clear thinking about the operational need for quick, automated localizations that can feed multi-messenger alerts. The soft spot is exactly where the stress-test note flags it. The abstract states that comparisons to chi-squared and maximum-likelihood methods are left for future work, and it gives no accuracy numbers, no coverage frequency for the reported 90% regions, no description of the simulation setup, and no checks on held-out real events or flight-like background. Without those, the claim that the network will produce well-calibrated regions on actual flight data remains untested. Generalization from idealized training counts to real trigger statistics and background variability is the load-bearing assumption, and nothing in the available text shows it holds. This is a paper for the COSI team and for groups building similar shield-based localizers. A reader already working on COSI data analysis or on machine-learning sky maps for transients would get the most from it, mainly as an architecture sketch they can try to reproduce and extend. It deserves a serious referee because the problem is mission-relevant and the technical framing is sound, even though the current version is preliminary. I would send it to review with the expectation that the authors add concrete validation and comparison sections before acceptance.

Circularity Check

0 steps flagged

No circularity: empirical DL method with no self-referential derivations

full rationale

The paper presents a neural network classifier that maps ACS shield counts to a HEALPix sky probability distribution for short GRB localization and 90% confidence region extraction. No equations, parameter fits, or derivations are shown that reduce the output to the input by construction. The work is framed as a proposed empirical technique trained on simulated data, with explicit deferral of validation against classical methods to future work. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps. The central claim therefore remains independent of its own fitted values or prior author results.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the neural-network weights themselves function as implicit fitted parameters whose values are not reported.

pith-pipeline@v0.9.0 · 5623 in / 1164 out tokens · 33849 ms · 2026-05-10T10:04:24.334179+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 4 canonical work pages

[1]

2015, The Astrophysical Journal Supplement Series, 216,

Connaughton, V ., et al. 2015, The Astrophysical Journal Supplement Series, 216,

2015
[2]

2025, bctools

URL https://dx.doi.org/10.1088/0067-0049/216/2/32 Martinez-Castellanos, I., et al. 2025, bctools. URLhttps://doi.org/10.5281/zenodo. 15419961 Parmiggiani, N., et al. 2026, The Astrophysical Journal, 997,

work page doi:10.1088/0067-0049/216/2/32 2025
[3]

2024, in 38th International Cosmic Ray Conference,

URLhttps://doi.org/ 10.3847/1538-4357/ae25fc Tomsick, J., et al. 2024, in 38th International Cosmic Ray Conference,

work page doi:10.3847/1538-4357/ae25fc 2024
[4]

Tomsick, S

eprint: 2308.12362 Zoglauer, A., et al. 2006, New Astronomy Reviews, 50,

work page arXiv 2006
[5]

V , URLhttps://doi.org/10.1016/j.newar.2006.06.049

Astronomy with Radioactivities. V , URLhttps://doi.org/10.1016/j.newar.2006.06.049

work page doi:10.1016/j.newar.2006.06.049 2006