HI Shells: Automatic Detection of Interstellar Shell Structures in Radio Data Cubes Using Machine Learning

HI Shells is an open-source machine learning project aimed at automatically identifying interstellar neutral hydrogen (HI) shell structures from 3D radio data cubes in FITS format. This project addresses the problem that traditional manual visual inspection is time-consuming and unable to handle massive datasets, providing an efficient automated tool for astrophysics research and helping to understand key scientific issues such as interstellar medium evolution and stellar feedback.

The structural evolution of the interstellar medium (ISM) is a core issue in astrophysics. As the main component of ISM, the distribution pattern of neutral hydrogen (HI) reveals processes such as star formation and supernova explosions. HI shells are expanding structures formed by stellar winds or supernova explosions, which are crucial for studying energy transfer mechanisms. Traditional manual detection is time-consuming and prone to missing small-scale/low signal-to-noise ratio structures. With the generation of terabyte-scale radio survey data (e.g., HI4PI, GALFA-HI), automated tools have become an urgent need.

HI Shells was developed by Ericturnip and hosted on GitHub. It is an open-source machine learning project specifically optimized for 3D radio data cubes in FITS format, capable of processing real radio survey observation data. It provides a complete code implementation for radio astronomers and demonstrates the application of deep learning in astrophysical data analysis.

Data Preprocessing: Remove baseline drift, correct uneven frequency response, convert intensity data format, and handle noise and radio frequency interference (RFI). Shell Feature Modeling: Learn the 3D features of HI shells—spatially, low-density regions are surrounded by high-density shell walls; in the spectral dimension, there are characteristic absorption/emission profiles. Machine Learning Architecture: Inference uses a 3D Convolutional Neural Network (3D CNN) to capture spatial-spectral correlation features and distinguish real shells from noise. Output and Visualization: Output shell positions (right ascension/declination), radial velocity, angular diameter, and confidence level. Support catalog construction and multi-band cross-matching.

Galactic Structure Research: Track large-scale structures, map spiral arm distribution, study the vertical structure of the galactic disk, and analyze gas exchange. Stellar Feedback Mechanism: Analyze the statistical properties of shells (size/velocity distribution), constrain supernova explosion rates and stellar wind energy injection efficiency, and improve galaxy evolution models. Multi-band Joint Analysis: Associate with HII regions, molecular clouds, and young star clusters, provide a reliable data foundation, and construct a complete star formation scenario.

Current Challenges: Diverse shell morphologies, confusion of boundaries due to foreground/background projection effects, and large angular scale differences requiring multi-scale analysis. Improvement Directions: Introduce attention mechanisms to handle multi-scale features, use GAN data augmentation to expand samples, and develop interpretability tools. Data Fusion Trend: The data volume of next-generation radio telescopes (e.g., SKA) will grow exponentially, making automated tools essential; multi-band data fusion helps understand the complete life cycle of the interstellar medium.

HI Shells represents a typical application of machine learning in astrophysics. It introduces advanced computing technology into manual data analysis processes, improves efficiency, and opens up new possibilities for interstellar medium evolution research. It is an open-source project worth attention in the interdisciplinary field of astrophysics and AI.