N-body simulations meet ML for structure detection and analysis
Recent advances in memory and processing power on small-scale computers enable new algorithms tailored to their capabilities. In astrophysics, two key tools benefit greatly: N-body simulations, which model systems of many interacting bodies and are used to study large-scale structure formation, test dark matter models, and analyze galaxy mergers and halos; and structure detection, which identifies galaxies, dark matter halos, filaments, walls, voids, and star-forming regions within the cosmic web.
In his thesis, Felipe Contreras Sepulveda introduces QUANTIZER, a new N-body simulation designed specifically for small-scale devices. Additionally, he enhances 1-DREAM, a toolkit of machine-learning algorithms. These updates include an innovative core design for the N-body code and ideas drawn from swarm intelligence, evolutionary computation, and probabilistic modeling. As a result, 1-DREAM more effectively detects, extracts, and models one-dimensional structures in complex datasets.
Contreras Sepulveda tested these methods on simulations of the cosmic web, galaxy collisions, synthetic datasets, and other systems, comparing their performance with established tools to evaluate accuracy and efficiency across varied scenarios.
After strengthening both methodologies, Contreras Sepulveda unified them into a single framework. QUANTIZER was used to simulate galaxy collisions, while 1-DREAM analyzed the resulting stellar streams and their evolution. This combined approach enables the study of astrophysical structures not only in space but also through time, offering a clearer and more dynamic view of their development. It reveals subtle patterns previously inaccessible and supports a deeper understanding of complex cosmic behavior.