Predicting far-infrared maps of galaxies via machine learning techniques

Context. The ultraviolet (UV) to sub-millimetre spectral energy distribution of galaxies can be roughly divided into two sections: the stellar emission (attenuated by dust) at UV to near-infrared wavelengths and dust emission at longer wavelengths. In Dobbels et al. (2020, A&A, 634, A57), we show that these two sections are strongly related, and we can predict the global dust properties from the integrated UV to mid-infrared emission with the help of machine learning techniques. Aims: We investigate if these machine learning techniques can also be extended to resolved scales. Our aim is to predict resolved maps of the specific dust luminosity, specific dust mass, and dust temperature starting from a set of surface brightness images from UV to mid-infrared wavelengths. Methods: We used a selection of nearby galaxies retrieved from the DustPedia sample, in addition to M31 and M33. These were convolved and resampled to a range of pixel sizes, ranging from 150 pc to 3 kpc. We trained a random forest model which considers each pixel individually. Results: We find that the predictions work well on resolved scales, with the dust mass and temperature having a similar root mean square error as on global scales (0.32 dex and 3.15 K on 18″ scales respectively), and the dust luminosity being noticeably better (0.11 dex). We find no significant dependence on the pixel scale. Predictions on individual galaxies can be biased, and we find that about two-thirds of the scatter can be attributed to scatter between galaxies (rather than within galaxies). Conclusions: A machine learning approach can be used to create dust maps, with its resolution being only limited to the input bands, thus achieving a higher resolution than Herschel. These dust maps can be used to improve global estimates of dust properties, they can lead to a better estimate of dust attenuation, and they can be used as a constraint on cosmological simulations that trace dust