Deep Learning for Position Angle Quantification Applied to Interstellar Filaments

Studying filamentary structures opened a door for understanding the conditions of star formation and interstellar medium properties. Present-day algorithms for identifying filaments and their orientations require careful individual parameterization for each astronomical map. The increasing scale of astronomical surveys presents significant challenges in handling larger data sets. While previous data resources have been sufficient, there is a growing demand for more extensive data sets to support advanced research. The construction of new astronomical facilities will substantially increase the volume of data available for exploration. For example, the Square Kilometre Array—the largest radio telescope project to date—will enable detailed mapping of the large-scale structure of the cosmos, including filaments, so the astronomical community is expected to process enormous amounts of data. In this study, we show the efficiency of a parameterization-free approach based on the U-Net convolutional neural network. We trained our model on a synthetically generated data set and tested it on observational data. Our method achieves a validation mean squared error of 7.73, demonstrating improved accuracy over existing approaches. The modular separation of preprocessing and angle determination makes it adaptable to different data sets. The variety of filamentary structures in the synthetic data set allows us to reproduce the multifariousness encountered in observational data. The resulting open-access model provides an efficient solution for position angle determination and filamentary structures identification that can be applied to large data sets.