Reimagining SED Fitting with Cosmological Galaxy Simulations and Machine Learning
SED fitting is the most common technique to recover galaxy physical properties from observed photometry. However, SED fitting requires many assumptions that essentially collapse a galaxy from a three-dimensional spatially varying object with complex structure into a scalar point. Moreover, modern inference techniques are computationally intensive, which presents a unique challenge in the era of extremely large datasets. We present \textsc{Phot-Gal}, a new galaxy SED modeling tool that solves the inverse problem of SED fitting by training a machine learning model on photometry generated from 3D radiative transfer of simulated galaxies with a wide range of implemented physics. \textsc{Phot-Gal} is designed to accept an arbitrary amount of input photometry by utilizing a $K$-nearest neighbors imputation strategy. Our fiducial model predicts redshift, stellar mass, dust mass, and star formation rate with uncertainties based on the provided input photometry. We evaluate the performance of \textsc{Phot-Gal} relative to the commonly-used SED fitting tool \textsc{prospector} in successfully recovering each of these properties with several metrics for the inferred values and uncertainties and find that it outperforms the accuracy of standard SED fitting software on the testing set. However, with fewer photometric constraints, \textsc{Phot-Gal} is more likely to have output uncertainties that do not reflect the offset from the ground truth. We dissect the components of \textsc{Phot-Gal} to find reasonable physical justifications for the photometry it relies on most, understand how each step in its workflow contributes to the eventual output posterior, and evaluate its ability to generalize to novel data.