Abstract

On large scales, the Universe forms a complex structure of filaments, walls, voids, and nodes known as cosmic web. It is known that a galaxy's location in the cosmic web can change their properties and affect their evolutionary history. However, it can be difficult to detect these structures observationally, since filaments are sparsely traced out by the few galaxies that are visible along their length. We propose to train a CNN-based encoder-decoder model on 3D cosmological simulations where the true filamentary structure is known, but only a small number of the galaxies that trace it out can be detected. Preliminary results have been achieved by combining multiple deep learning techniques such as kernel design, pre-training and data augmentation. This has led to an accuracy of 97.6%, precision of 33.0% and recall of 41.4%, which in the context of the problem means success in identifying true negatives and room for improvement in true positives.