The formation of stars in galaxies occurs in filaments made up of gas, primarily hydrogen, and small solid particles called interstellar dust, primarily composed of carbon. These filaments can be difficult to detect in data depending on their location and physical properties, such as density and temperature. In particular, low-density filaments or those located in areas of strong emission are often not detected.

A team of researchers, including those from CNRS laboratories, have implemented an innovative and interdisciplinary approach using supervised machine learning to try and detect filaments located in the plane of our galaxy. This approach builds on existing results from traditional filament detection methods. The extracted filaments are used to train Unet and Unet++ type convolutional neural networks.

The trained model learns to recognize the filaments and creates an image of the galactic plane where each pixel is represented by its probability, ranging from 0 to 1, of belonging to the “filament” class.

The results of the machine learning approach demonstrate that this method can detect filaments that were not previously identified using traditional detection methods. These new filaments can be confirmed through an empirical approach using data at other wavelengths that are currently not used in the learning process. The goal of this BigSF project, which is funded by the CNRS’s mission for transverse and interdisciplinary initiatives (MITI), is to study the formation of stars in our galaxy by utilizing a large amount of available data and automatic learning techniques.