Explaining Dark Matter Halo Density Profiles with Neural Networks.

Area of Science:

• Astrophysics
• Cosmology
• Machine Learning

Background:

• Understanding dark matter halo formation and evolution is crucial in cosmology.
• Density profiles of dark matter halos encode information about their assembly history.
• Traditional methods struggle to fully disentangle complex relationships in large astrophysical datasets.

Purpose of the Study:

• To connect the evolutionary history of dark matter halos with their density profiles using explainable neural networks.
• To identify and interpret the independent factors of variation within halo density profiles.
• To explore the potential of machine learning for scientific discovery in astrophysics.

Main Methods:

• Utilized explainable neural networks (XNNs) to analyze dark matter halo data.
• Employed a low-dimensional representation to capture key variations in density profiles.
• Applied mutual information to physically interpret the learned representations.

Main Results:

• The XNN successfully recovered known relationships between early halo assembly and inner density profiles.
• A novel discovery identified a single parameter, related to recent mass accretion rate, that describes the halo profile beyond the virial radius.
• The network learned interpretable features without prior evolutionary knowledge.

Conclusions:

• Explainable neural networks offer a powerful tool for astrophysical data analysis.
• Machine learning can facilitate scientific discovery by uncovering hidden relationships in complex datasets.
• The study provides new insights into the factors governing dark matter halo density profiles.