Traditional cosmic void-finding algorithms require extensive computational resources, severely limiting their applicability to large-scale cosmological surveys. We present a novel deep learning approach using a 3D adaptation of the YOLO-like object detection architecture that reduces this computational burden by orders of magnitude, while maintaining a very good detection accuracy. Our method processes voxelised particle density fields at 2 h − 1 Mpc resolution using a Feature Pyramid Network architecture to detect voids at different size ranges simultaneously. In particular, we focus on scales where traditional void-finders are computationally less efficient due to the sheer abundance of structures in the Universe, i.e. for void-radii in the range of 10 ≲ R ≲ 35 h − 1 Mpc . This is an intrinsic limitation of our solution; nevertheless, for radii R ≳ 35 h − 1 Mpc traditional spherical void-finders are very efficient, therefore accelerated approaches are not required. In this work, we evaluate our method on comoving snapshots of cosmological N-body simulations, achieving 85% precision and 89% recall, an average spherical Intersection over Union of 47%, and an optimal performance with 87% F1-score, for voids in the above radius range . Moreover, we also recover within errors void summary statistics, such as the void size function. Our approach constitutes a trade-off different from others in the literature, able to maintain a sufficient accuracy compared to traditional spherical void-finding methods, with a high gain in computational efficiency. This work represents one of the first applications of modern object detection architectures to 3D cosmological voids structure identification, enabling real-time void analysis for large-scale surveys and comprehensive cosmological parameter studies.
Puglisi et al. (Mon,) studied this question.