Fingerprint matching of beyond-WIMP dark matter: Neural network approach

Improving observation of galactic-scale structure provide important clues to dark matter properties. While weakly interacting massive particles (WIMPs) provide cold dark matter on galactic scales, beyond-WIMP candidates suppress the galactic-scale structure formation. Nevertheless, directly constraining microscopic model parameters from observations involves an interdisciplinary and time-consuming procedure. In practice, some parametrizations of the linear matter power spectrum are introduced. The particle physics community calculates the linear matter power spectrum for a given model parameter set, while the astrophysics community places the constraint on the power spectrum parameters. If maps between the model parameters and the power spectrum parameters and maps between the power spectrum parameters and the likelihood (or observables) are shared among the two communities, they are very useful for both communities, e.g., making a constraint plot of the model parameter space. As suggested in the literature, however, it is necessary to introduce multiple parameters to precisely describe the linear matter power spectrum in a wide range of beyond-WIMP models. It challenges us to express and share the non-linear maps between multiple parameters. In this work, we propose utilizing the neural network technique to this end. The neural network technique is known to automatically express and efficiently share non-linear maps, although it is not as simple as analytic fitting formulas if available. To demonstrate how to work with a concrete example, we consider a simplified model of light feebly interacting massive particles and simple observables for galactic-scale structure. We also reveal the obtained neural networks through the arXiv website.