Fast and Accurate Stellar Mass Predictions from Broad-Band Magnitudes with a Simple Neural Network: Application to Simulated Star-Forming Galaxies

TL;DR

This study demonstrates that a simple neural network trained on broad-band photometry can rapidly and accurately estimate stellar masses of star-forming galaxies, outperforming traditional methods in efficiency and ease of use.

Contribution

The paper introduces a straightforward neural network model that effectively predicts stellar masses from photometric data, showing high accuracy without complex physical modeling.

Findings

Achieves 0.085 dex RMS error in stellar mass predictions

Identifies FUV - NUV colour as a key predictor

Provides a fast, data-driven alternative to SED fitting

Abstract

A simple, fully connected neural network with a single hidden layer is used to estimate stellar masses for star-forming galaxies. The model is trained on broad-band photometry - from far-ultraviolet to mid-infrared wavelengths - generated by the Semi-Analytic Model of galaxy formation (SHARK), along with derived colour indices. It accurately reproduces the known SHARK stellar masses with respective root-mean-square and median errors of only 0.085 and 0.1 dex over the three decades in stellar mass. Analysis of the trained network's parameters reveals several colour indices to be particularly effective predictors of stellar mass. In particular, the FUV - NUV colour emerges as a strong determinant, suggesting that the network has implicitly learned to account for attenuation effects in the ultraviolet bands, thereby increasing the diagnostic power of this index. Traditional methods such as spectral energy distribution fitting, though widely used, are often complex, computationally expensive, and sensitive to model assumptions and parameter degeneracies. In contrast, the neural network relies solely on easily obtained observables, enabling rapid and accurate stellar mass predictions at minimal computational cost. The model derives its predictions exclusively from patterns learned in the data, without any built-in physical assumptions (such as stellar initial mass function). These results demonstrate the utility of this study's machine learning approach in astrophysical parameter estimation and highlight its potential to complement conventional techniques in upcoming large galaxy surveys.