发表机构
Texas State University（德克萨斯州立大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
XClass是一个自动化多波段机器学习流水线，通过SED转换和两阶段随机森林，将河外星系X射线源分为七类，在11,374个源上达到99.6%准确率。
AI 中文摘要
钱德拉X射线天文台在近邻星系中探测到的大部分X射线源缺乏天体物理分类。我们提出了XClass（河外星系源X射线分类器），这是一个端到端的机器学习流水线，将河外星系X射线点源分为七类：活动星系核（AGN）、低质量X射线双星（LMXBs）、高质量X射线双星（HMXBs）、猫状变星（CVs）、低质量和高质量前景恒星，以及超新星遗迹。关键挑战在于测光异质性：训练源主要来自银河系，其测光数据来自广域巡天（PanSTARRS和2MASS），而河外星系目标则需要哈勃空间望远镜（HST）在不相交的滤光片系统中的成像。我们通过光谱能量分布（SED）转换来解决这一问题，该转换将适合各类别的光谱模型拟合到每个训练源，并将最佳拟合模型通过HST滤光片曲线进行卷积，从而在共同的特征空间中产生合成星等。该分类器采用非对称的两阶段随机森林：第一阶段区分宽泛类别（AGN、X射线双星、超新星遗迹、恒星），第二阶段使用包含第一阶段概率的增强特征向量将X射线双星区分为LMXBs和HMXBs。训练集由十个银河系星表和外星系超新星遗迹星表组装而成，并与钱德拉源星表v2.1交叉匹配。特征包括X射线硬度比、SED转换的HST颜色以及X射线到光学通量比。我们将训练集限制为至少有一个光学星等的源，以避免插补伪影；该流水线在由此产生的11,374个源的光学基线上实现了99.6%的准确率和0.90的平衡准确率，且校准极佳（ECE = 0.002）。XClass是模块化的，可推广到任何HST滤光片配置，并将在配套论文中应用于M31和M33。
英文摘要
Most X-ray sources detected in nearby galaxies by the Chandra X-ray Observatory lack astrophysical classifications. We present XClass (X-ray Classifier for extragalactic sources), an end-to-end machine-learning pipeline that classifies extragalactic X-ray point sources into seven classes: AGN, LMXBs, HMXBs, CVs, low-mass and high-mass foreground stars, and supernova remnants. The key challenge is photometric heterogeneity: training sources are predominantly Galactic with photometry from wide-field surveys (PanSTARRS and 2MASS), while extragalactic targets require Hubble Space Telescope (HST) imaging in a disjoint filter system. We address this through a spectral energy distribution (SED) translation that fits class-appropriate spectral models to each training source and convolves the best-fit model through HST filter curves, producing synthetic magnitudes in a common feature space. The classifier uses an asymmetric two-stage Random Forest: Stage 1 separates broad categories (AGN, X-ray binaries, SNRs, stars) and Stage 2 resolves X-ray binaries into LMXBs and HMXBs using an augmented feature vector that includes Stage 1 probabilities. The training set is assembled from ten Galactic catalogs and extragalactic SNR catalogs, cross-matched with the Chandra Source Catalog v2.1. Features include X-ray hardness ratios, SED-translated HST colors, and X-ray-to-optical flux ratios. We restrict the training set to sources with at least one optical magnitude, avoiding imputation artifacts; the pipeline achieves 99.6% accuracy and balanced accuracy of 0.90 on the resulting 11,374-source optical baseline, with excellent calibration (ECE = 0.002). XClass is modular, generalizable to any HST filter configuration, and will be applied to M31 and M33 in a companion paper.
Comments17 pages, 7 figures. Published in ApJ
Journal refAstrophys. J. 1008, 117 (2026)
