Galactic-scale structures can play an important role in pinning down dark matter properties. While weakly interacting massive particles (WIMPs) behave as cold dark matter on galactic scales, it is known that many beyond-WIMP candidates suppress the linear matter power spectrum. Though the suppression has been traditionally parametrized by a single parameter, thermal warm dark matter mass, the actual suppression in the matter power depends on the underlying mechanism and is much more complicated. Thus, in order to prepare for future observational improvement, it is necessary to (1) introduce a parameter set that covers a wide range of beyond-WIMP models, and (2) develop a method to connect such parameters with the underlying model parameters and/or observables in an efficient manner. In this talk we propose using neural network technique for the latter purpose. We demonstrate how the connection is realized in a ready-to-use manner by neural network, taking light feebly interacting massive particles (FIMPs) as an example.