The rapid release of accurate sky localization for gravitational-wave (GW) candidates is crucial for multi-messenger observations. During the third observing run of Advanced LIGO and Advanced Virgo, automated GW alerts were publicly released within minutes of detection. Subsequent inspection and analysis resulted in the eventual retraction of a fraction of the candidates. Updates could be delayed by up to several days, sometimes issued during or after exhaustive multi-messenger follow-up campaigns. We introduce GWSkyNet, a real-time framework to distinguish between astrophysical events and instrumental artifacts using only publicly available information from the LIGO-Virgo open public alerts. This framework consists of a non-sequential convolutional neural network involving sky maps and metadata. GWSkyNet achieves a prediction accuracy of 93.5% on a testing data set.

快速释放重力波（GW）候选人的准确天空定位对于多信使观测至关重要。在Advanced LIGO和Advanced Virgo的第三次观测运行期间，自动GW警报在检测到的几分钟内公开发布。随后的检查和分析导致部分候选人最终退缩。更新可能会延迟几天，有时会在详尽的多信使后续活动期间或之后发布。我们引入GWSkyNet，这是一个实时框架，仅使用LIGO-Virgo开放式公共警报中的公共可用信息来区分天体物理事件和器物。该框架由涉及天空图和元数据的非顺序卷积神经网络组成。GWSkyNet在测试数据集上的预测准确性达到93.5％。