Interpolated kilonova spectra models: necessity for a phenomenological, blue component in the fitting of AT2017gfo spectra

In this work, we present a simple interpolation methodology for spectroscopic time series, based on conventional interpolation techniques (random forests) implemented in widely-available libraries. We demonstrate that our existing library of simulations is sufficient for training, producing interpolated spectra that respond sensitively to varied ejecta parameter, post-merger time, and viewing angle inputs. We compare our interpolated spectra to the AT2017gfo spectral data, and find parameters similar to our previous inferences using broadband light curves. However, the spectral observations have significant systematic short-wavelength residuals relative to our models, which we cannot explain within our existing framework. Similar to previous studies, we argue that an additional blue component is required. We consider a radioactive heating source as a third component characterized by light, slow-moving, lanthanide-free ejecta with $M_{\rm th} = 0.003~M_\odot$, $v_{\rm th} = 0.05$c, and $\kappa_{\rm th} = 1$ cm$^2$/g. When included as part of our radiative transfer simulations, our choice of third component reprocesses blue photons into lower energies, having the opposite effect and further accentuating the blue-underluminosity disparity in our simulations. As such, we are unable to overcome short-wavelength deficits at later times using an additional radioactive heating component, indicating the need for a more sophisticated modeling treatment.