Pattern Finding in mm-Wave Spectra of Massive Young Stellar Objects

TL;DR

This study uses advanced dimensionality reduction and machine learning techniques on ALMA spectra to classify massive young stellar objects into evolutionary groups based on their chemical and physical properties.

Contribution

It introduces a novel approach combining Locally Linear Embedding, Gaussian Mixture Models, PCA, and random forests to analyze MYSO spectra without manual line fitting.

Findings

Identified three distinct MYSO groups with different chemical and physical conditions.

Demonstrated that spectra can predict physicochemical states using machine learning.

Revealed an evolutionary sequence among MYSOs through spectral analysis.

Abstract

Massive stars play a pivotal role in shaping their galactic surroundings due to their high luminosity and intense ionizing radiation. However, the precise mechanisms governing the formation of massive stars remain elusive. Complex organic molecules (COMs) offer an avenue for studying star formation across the low- to high-mass spectrum because COMs are found in every young stellar object phase and offer insight into the structure and temperature. We aim to unveil evolutionary patterns of COM chemistry in 41 massive young stellar objects (MYSOs) sourced from diverse catalogues, using ALMA Band 6 spectra. Previous line analysis of these sources showed the presence of CH$_3$OH, CH$_3$CN, and CH$_3$CCH with diverse excitation temperatures and column densities, indicating a possible evolutionary path across sources. However, this analysis usually involves manual line extraction and rotational diagram fitting. Here, we improve upon this process by directly retrieving the physicochemical state of MYSOs from their dimensionally-reduced spectra. We use a Locally Linear Embedding to find a lower-dimensional projection for the physicochemical parameters obtained from individual line analysis. We identify clusters of similar MYSOs in this embedded space using a Gaussian Mixture Model. We find three groups of MYSOs with distinct physicochemical conditions: i) cold, COM-poor sources, ii) warm, medium-COM-abundance sources, and iii) hot, COM-rich sources. We then apply principal component analysis (PCA) to the spectral sample, finding further evidence for an evolutionary path across MYSO groups. Finally, we find that the physicochemical state of our sample can be derived directly from the spectra by training a simple random forest model on the first few PCA components. Our results highlight the effectiveness of dimensionality reduction in obtaining clear physical insights directly from MYSO spectra.