利用卷积神经网络搜索暗能量光谱仪模拟光谱中的莱曼极限系统  

Searching for Lyman limit systems in Dark Energy Spectroscopic Instrument mock spectra using convolutional neural network

作　　者：蒲巧 刘涵 邹佳琪 蔡峥 Qiao Pu;Han Liu;Jiaqi Zou;Zheng Cai(School of Computer Technology and Applications,Qinghai University,Xining 810016,China;Department of Astronomy,Tsinghua University,Beijing 100084,China)

机构地区：[1]青海大学计算机技术与应用学院,西宁810016 [2]清华大学天文系,北京100084

出　　处：《科学通报》2024年第21期3191-3198,共8页Chinese Science Bulletin

基　　金：国家自然科学基金(12073014)资助。

摘　　要：研究莱曼极限系统(Lyman limit systems,LLS)对于了解宇宙的大尺度结构、星系演化以及星系团内部气体分布具有重要意义.然而,由于LLS吸收特征的独特性,目前的研究主要采用传统方法,对柱密度在10^(19)cm^(-2)≤N(HI)<10^(20.3)cm^(-2)的小样本集上进行认证.本文利用深度学习技术,在暗能量光谱仪(The Dark Energy Spectroscopic Instrument,DESI)模拟光谱上,通过优化卷积神经网络(convolutional neural network,CNN)模型,提高了对LLS(10^(18.5)cm^(-2)≤N(HI)≤10^(20.0)cm^(-2))的识别精度(达到95%).随后,验证了该模型的完备度和纯度,并估计了LLS的柱密度和红移.结果显示:在S/N>6的情况下,当10^(19.0)cm^(-2)>N(HI)>1018.5cm^(-2)时,CNN模型的完备度超过0.5,而纯度也超过0.2;当10^(20.0)cm^(-2)>N(HI)>10^(19.0)cm^(-2)时,完备度超过0.9,而纯度超过0.7;当10^(20.0)cm^(-2)>N(HI)>10^(18.5)cm^(-2)时,CNN模型对LLS柱密度估计值与真实值的平均差值为-0.05161,标准差为0.239,对LLS红移估计值和真实值的平均差值为-0.0003,标准差为0.0009.这些结果表明:模型的完备度普遍高于纯度,尤其是在低柱密度的情况下,LLS在光谱中的吸收特征非常窄,极易与其他波段混淆,导致模型产生更多的FP(false positive)样本.此外,CNN模型对LLS的柱密度和红移的估计值略低于真实值,且估计误差的离散程度较小.本研究为未来的LLS研究提供了可参考的方法,鼓励研究人员适应并采用CNN模型进行各种光谱分析.Studying Lyman limit systems(LLS)is crucial for a deeper understanding of the large-scale structure of the universe,the evolutionary history of galaxies,and the distribution of gas within galaxy clusters.Although LLS absorption features are distinctive,current research is largely constrained by these characteristics.Additionally,traditional methods are predominantly employed,with a primary focus on the identification and analysis of small sample sets with column densities ranging from 10^(19)cm^(−2)≤N(HI)<10^(20.3)cm^(−2).The objective of this study is to surpass the constraints of current research by utilizing deep learning methods to investigate a wider and more inclusive sample.This approach facilitates the detection and characterization of LLS with reduced column densities.We utilized high-quality spectral data simulated by the Dark Energy Spectroscopic Instrument(DESI)as the experimental foundation.Through the optimization of convolutional neural network(CNN)models,we have effectively boosted the model’s identification accuracy of LLS(with column densities of 1018.5 cm^(−2)≤N(HI)≤10^(20.0)cm^(−2))in DESI simulated spectra to 95%.Following that,this paper validated the completeness and purity of the model under different signal-to-noise ratios and column density conditions.Additionally,an analysis of the differences between the CNN model’s estimated and actual values of column density and redshift was conducted.The analysis results indicate that,under conditions where the signal-to-noise ratio exceeds 6,for LLS with column densities of 10^(19.0)cm^(−2)>N(HI)>1018.5 cm^(−2),the completeness of the CNN model exceeds 0.5,and the purity exceeds 0.2.For LLS with column densities of 10^(20.0)cm^(−2)>N(HI)>10^(19.0)cm^(−2),the model’s completeness exceeds 0.9,and the purity exceeds 0.7.Further analysis reveals that as column density and signal-to-noise ratio increase,both the completeness and purity of the model exhibit an upward trend.In the comparison between the estimated and actual va

关 键 词：DESI类星体光谱 莱曼极限系统 卷积神经网络 完备度和纯度 

分 类 号：P111[天文地球—天文学] P158