深度學習於星系團成員之應用

星系團成員星系的判斷對於星系演化、星系團質量和宇宙學等研究至關重要。在過去的二十年裡，已經有好幾種星系團成員星系的判斷方法被開發了。一般來說，有三種方法，第一是基於星系顏色與亮度的方法，例如紅序列(red sequence)；第二是基於紅移的方法，研究人員透過直接測量星系的光度紅移(photometric redshift, photo-z)或光譜紅移(spectroscopic redshift, spec-z)來判斷我們與該星系的距離，第三則是基於機器學習(machine learning, ML)或深度學習(deep learning, DL)，直接進行星系團成員星系的判斷。近年來，基於機器學習或深度學習的方法為光度紅移及星系團成員星系的判斷帶來更高效率且更好的結果。但是，這些研究都是基於大量的光譜能量分布(spectral energy distribution, SED)的資訊，也就是說，多波段，這些研究人員通常使用五個以上的波段。在我們的研究當中，我們想要知道，利用兩個波段及非SED的資訊，例如星系的表面亮度或是形狀，是否能夠得到與其他紅移估計與星系團成員辨認相關研究相當，或是更好的結果，同時，我們也設置了一系列的深度學習實驗來了解怎樣的來源，前景或是背景星系，會對星系團成員星系的辨認造成影響。我們的研究結果顯示，使用兩個波段及非SED的資訊在紅移估計上得到可與其他研究可相提並論之結果，我們模型的均方根誤差(root mean squared error)大約為0.08，而平均絕對誤差(mean absolute error)大約為0.06，且光學波段(V band)對於紅移的估計相對重要。在星系團成員星系的判斷上，我們得到70 %的ROC下面積(area under receiver operating characteristic curve, AUC)，前景星系對於星系團成員星系的判斷會造成問題，以及利用不同視線上速率(line-of-sight velocity)來限制星系團成員星系的範圍並不會對結果產生影響。除此之外，我們透過比較利用深度學習以及利用預測的紅移，這兩種方式在星系團成員星系的判斷上得知，利用預測的紅移來判斷星系團成員星系是不可能的，因為預測紅移的模型誤差比星系團的紅移範圍還要大。在深度學習模型方面，我們發現到利用適當數量的資料訓練多層感知器與卷積神經網路的混和模型(hybrid MLP-CNN model)，通常能夠得到較好且較穩定的結果，這樣的結果顯示讓深度學習模型同時學習物體的特徵數值及結構是較好的訓練策略。

關鍵字

深度學習；星系團

並列摘要

Galaxy cluster membership assignment is crucial to the researches related to galaxy evolution, the mass of galaxy cluster, and cosmology. Over the past twenty years, several methods for galaxy cluster membership assignment have been developed. In general, there are three kinds of methods, first, color-magnitude-based method, like red sequence, second, redshift-based methods, researchers directly measured the distance between us and the galaxies by photometric redshift (photo-z) or spectroscopic redshift (spec-z), third, the ML/DL-based methods, researchers used machine learning (ML) or deep learning (DL) to do galaxy cluster membership assignment directly. In recent years, ML/DL-based methods bring more efficient and even or higher performance to the photo-z and galaxy cluster membership assignment. However, all of them are based on the enormous amount of SED information, i.e. multi-bands, typically, those researchers used >5 bands. Our study wants to know whether using state-of-the-art deep learning models trained by two bands with non-SED information, for example, surface brightness and the shape of the galaxies will bring us even or higher performance in redshift estimation and galaxy cluster membership assignment or not. Also, we set up serial deep learning experiments to understand what kind of sources, foreground or background, will result in a problem when we do galaxy cluster membership assignment. Our results reveal that using two bands with non-SED information can achieve comparable results in redshift estimation. We got root mean squared error ~0.08 and mean absolute error ~0.06, and V band is vital for redshift estimation. Our best results for galaxy cluster membership assignment is ~70% area under receiver operating characteristic curve (AUC). We found that the foreground sources cause a problem in galaxy cluster membership assignment, and using different line-of-sight velocity to define the galaxy cluster members will not affect the results. Besides, we compare directly DL membership to inferred membership based on predicted redshift, and we found that using predicted redshift to do galaxy cluster membership assignment is impossible because the error of the model of redshift estimation is larger than the redshift range of a galaxy cluster. When it comes to the deep learning models, hybrid MLP-CNN model trained with an appropriate amount of data usually achieve higher and more stable results, which means that let deep learning models learn the features and structure simultaneously is a better training strategy.