In this paper we focus on distances, as there is a ground truth measurement of comparison. But there are some early signs that we are getting reasonable estimates for, e.g., stellar mass here as well (below: an injection/recovery test placing a logM=7.4 galaxy at many distances).

The method actually appears to be working rather well! Below is an example of one real input galaxy and three random draws from the posterior. On a sample of ~20 nearby galaxies with literature distances from TRGB/SBF, silkscreen almost always recovers the distance within the 5-95% posterior.

To do this we created Silkscreen, a code which uses neural posterior estimation and rounds of simulation and training to fit a set of galaxy properties given input images. In short, we simulate galaxy images with artpop, and use a ResNet to extract summary statistics before passing to the sbi code.

Our idea stems from the ability for the Artpop code to make realistic galaxy images from first principles -- i.e., create stellar populations sampling IMFs and place them "star by star" in some distribution (e.g., Sérsic), while also simulating survey depths and noise properties.

Thrilled to announce an event I've long been thinking about! Join us at Yale University for a "Hack Day Conference," in which a group of educators come together to build resources and exercises for teaching research methods in astronomy. See astrocodex.github.io for more.

After half a decade of getting to use this incredible facility, today I got to visit Keck! It was both awe inspiring to see the telescopes and instruments that have defined my time in astronomy, and eye opening to see how much work is needed to keep this place looking upward night after night.