Leveraging Machine Learning for Accurate and Fast Stellar Mass Estimation of Galaxies

Unveiling the evolutionary history of galaxies necessitates a precise understanding of their physical properties. Traditionally, astronomers achieve this through spectral energy distribution (SED) fitting. However, this approach can be computationally intensive and time-consuming, particularly for large datasets. This study investigates the viability of machine learning (ML) algorithms as an alternative to traditional SED-fitting for estimating stellar masses in galaxies. We compare a diverse range of unsupervised and supervised learning approaches including prominent algorithms such as K-means, HDBSCAN, Parametric t-Distributed Stochastic Neighbor Embedding (Pt-SNE), Principal Component Analysis (PCA), Random Forest, and Self-Organizing Maps (SOM) against the well-established LePhare code, which performs SED-fitting as a benchmark. We train various ML algorithms using simple model SEDs in photometric space, generated with the BC03 code. These trained algorithms are then employed to estimate the stellar masses of galaxies within a subset of the COSMOS survey dataset. The performance of these ML methods is subsequently evaluated and compared with the results obtained from LePhare, focusing on both accuracy and execution time. Our evaluation reveals that ML algorithms can achieve comparable accuracy to LePhare while offering significant speed advantages (1,000 to 100,000 times faster). K-means and HDBSCAN emerge as top performers among our selected ML algorithms. Supervised learning algorithms like Random Forest and manifold learning techniques such as Pt-SNE and SOM also show promising results. These findings suggest that ML algorithms hold significant promise as a viable alternative to traditional SED-fitting methods for estimating the stellar masses of galaxies.