Silkscreen: Direct Measurements of Galaxy Distances from Survey Image Cutouts

TL;DR

Silkscreen is a novel method combining neural posterior estimation and convolutional neural networks to directly infer galaxy distances from survey images, aiming to enable rapid analysis of large upcoming astronomical surveys.

Contribution

The paper introduces silkscreen, a new machine learning and Bayesian inference framework for estimating galaxy distances directly from images, improving applicability in the semi-resolved regime.

Findings

Accurately recovers distances for nearby galaxies using survey images.

Demonstrates potential for rapid inference in large-scale surveys like LSST.

Addresses limitations of current distance estimation techniques.

Abstract

With upcoming wide field surveys from the ground and space the number of known dwarf galaxies at $\lesssim 25$ Mpc is expected to dramatically increase. Insight into their nature and analyses of these systems' intrinsic properties will rely on reliable distance estimates. Currently employed techniques are limited in their widespread applicability, especially in the semi-resolved regime. In this work we turn to the rapidly growing field of simulation based inference to infer distances, and other physical properties, of dwarf galaxies directly from multi-band images. We introduce silkscreen: a code leveraging neural posterior estimation to infer the posterior distribution of parameters while simultaneously training a convolutional neural network such that inference is performed directly on the images. Utilizing this combination of machine learning and Bayesian inference, we demonstrate the method's ability to recover accurate distances from ground-based survey images for a set of nearby galaxies ($2 < D ({\rm Mpc)} < 12$) with measured SBF or TRGB distances. We discuss caveats of the current implementation along with future prospects, focusing on the goal of applying silkscreen to large upcoming surveys, like LSST. While the current implementation performs simulations and training on a per-galaxy basis, future implementations will aim to provide a broadly-trained model that can facilitate inference for new dwarf galaxies in a matter of seconds using only broadband cutouts. We focus here on dwarf galaxies, we note that this method can be generalized to more luminous systems as well.