Seminars : Documents

Dust and stellar property estimates via machine learning techniques
Wouter Dobbels (U. Gent)

Large surveys have been performed from the ultraviolet (UV) to the far-infrared (FIR). Some galaxies are observed over this whole wavelength range, and through SED fitting we can estimate the stellar and dust properties. Unfortunately, most galaxies are only detected at a limited part of this spectrum.

With machine learning techniques, we can use the UV-FIR galaxies as a blueprint: we learn the mapping from a subset of their fluxes to their properties. For example, a mapping from UV-NIR (i.e. stellar emission) to dust mass can be established, and then applied to galaxies that lack FIR data. Whereas traditional SED fitting methods can only estimate dust mass using FIR data, our method is accurate (RMSE = 0.3 dex) and unbiased in the absence of FIR. Besides what can be directly estimated from the SEDs alone, this technique implicitly uses relations that follow from galaxy evolution. To avoid a black box, we take special care to estimate uncertainties on our predictions and to interpret the model. The main idea can be generalised: a sample with a broad set of observations can be used as a blueprint, to improve estimates of samples with limited observations.