发表机构
Institute of Advanced Research; St.Xavier’s College (Autonomous); Space Research Institute, Austrian Academy of Sciences; Nikhef; Centre for Development of Telematics (C-DOT); Department of Physics, University of Massachusetts, Dartmouth（高级研究所; 圣泽维尔学院（自治）; 奥地利科学院空间研究所; 荷兰国家核与粒子物理研究所; 电信发展中心; 马萨诸塞大学达特茅斯分校物理系）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究将CNN参数空间约束框架扩展到对齐自旋双黑洞，证明模板库训练足够，且啁啾质量-持续时间表示在补丁识别中准确率最高（93.2%），显著提升引力波搜索效率。
AI 中文摘要
使用匹配滤波探测致密双星并合（CBCs）产生的引力波（GWs）在计算上要求很高，因为探测器数据必须与许多跨越高维内在参数空间的模板波形进行相关。在我们先前的工作中，我们展示了卷积神经网络（CNN）能够将含噪信号与纯噪声进行分类，并约束真实非自旋双黑洞（BBH）信号的内在参数空间，从而实现更窄的匹配滤波搜索区域并降低计算成本。在此，我们将该框架扩展到对齐自旋BBH系统，并研究不同的参数空间表示如何影响基于CNN的补丁识别。使用IMRPhenomD波形覆盖对齐自旋BBH参数空间以及先进LIGO设计灵敏度，我们表明，仅在对齐自旋模板库上训练的CNN在独立生成的均匀采样BBH信号上实现了超过99.8%的信号-噪声分类准确率，表明额外的均匀采样训练数据是不必要的。我们使用基于主成分分析（PCA）的分位数方案将模板库划分为四个近似平衡的补丁，并评估五种参数空间表示用于补丁识别。啁啾质量-持续时间表示实现了最高的平均准确率（93.2%），其次是啁啾质量（91.6%）和分量质量（90.5%）。后牛顿坐标tau_0-tau_3和theta_0-theta_3-theta_3s产生的准确率显著较低。这些结果表明，现有的模板库足以用于训练和高精度信号检测，而参数空间表示的选择对于约束真实信号的参数至关重要。
英文摘要
The detection of gravitational waves (GWs) from compact binary coalescences (CBCs) using matched filtering is computationally demanding because detector data must be correlated with many template waveforms spanning a high-dimensional intrinsic parameter space. In our previous work, we showed that a convolutional neural network (CNN) can classify noisy signals against pure noise and constrain the intrinsic parameter space of a true non-spinning binary black hole (BBH) signal, enabling a narrower matched-filter search region and reducing computational cost. Here, we extend this framework to aligned-spin BBH systems and investigate how different parameter-space representations affect CNN-based patch identification. Using IMRPhenomD waveforms over the aligned-spin BBH parameter space and Advanced LIGO design sensitivity, we show that a CNN trained solely on the aligned-spin template bank achieves more than 99.8% signal-noise classification accuracy on independently generated uniformly sampled BBH signals, indicating that additional uniformly sampled training data are unnecessary. We partition the template bank into four approximately balanced patches using a Principal Component Analysis (PCA)-based quantile scheme and evaluate five parameter-space representations for patch identification. The chirp mass-duration representation achieves the highest average accuracy (93.2%), followed by chirp mass (91.6%) and component masses (90.5%). The post-Newtonian coordinates tau_0-tau_3 and theta_0-theta_3-theta_3s yield substantially lower accuracies. These results show that the existing template bank is sufficient for training and high-accuracy signal detection, while the choice of parameter-space representation is critical for constraining the parameters of the true signal.
