AGNet: Weighing Black Holes with Deep Learning

TL;DR

AGNet introduces a CNN-MLP hybrid to weigh SMBHs directly from multi-band quasar light curves, bypassing expensive spectroscopy. By encoding Stripe 82 light curves into images and fusing learned representations with hand-crafted photometric features, it predicts SMBH masses and redshifts with competitive accuracy (SMBH mass RMSE ≈ $0.371$ dex and redshift RMSE ≈ $0.373$) against virial mass estimates. The approach demonstrates that photometric time-series contain sufficient information to estimate SMBH properties at a scale compatible with LSST, and the method shows clear benefits over purely CNN, MLP, or KNN baselines. These results pave the way for scalable SMBH demographics and rapid follow-up planning in upcoming surveys, with public code and clear avenues for uncertainty quantification and improved ground-truth calibration via reverberation mapping.

Abstract

Supermassive black holes (SMBHs) are ubiquitously found at the centers of most massive galaxies. Measuring SMBH mass is important for understanding the origin and evolution of SMBHs. However, traditional methods require spectroscopic data which is expensive to gather. We present an algorithm that weighs SMBHs using quasar light time series, circumventing the need for expensive spectra. We train, validate, and test neural networks that directly learn from the Sloan Digital Sky Survey (SDSS) Stripe 82 light curves for a sample of $38,939$ spectroscopically confirmed quasars to map out the nonlinear encoding between SMBH mass and multi-color optical light curves. We find a 1$σ$ scatter of 0.37 dex between the predicted SMBH mass and the fiducial virial mass estimate based on SDSS single-epoch spectra, which is comparable to the systematic uncertainty in the virial mass estimate. Our results have direct implications for more efficient applications with future observations from the Vera C. Rubin Observatory. Our code, \textsf{AGNet}, is publicly available at \url{https://github.com/snehjp2/AGNet}.

Figures (7)

• Figure 1: Distribution of $417,618$ cleaned DR14 quasars in mass-redshift space with top and right panels showing quasar redshift and SMBH mass histograms, respectively. Mass is measured in units of $log(M_{SMBH}/M_{\sun})$. Redshifts are provided in the DR7 and DR14 catalogs and cross-matched with redshift values from the Stripe 82 light curves. Unphysical supermassive black hole masses ($M_{SMBH}<0$) and unreliable estimates with error $>.3$ dex are removed.
• Figure 2: Convolutional Neural Network Architecture. A basic ResNet block is composed of two layers of 3x3 convolutions, subsequently applying a batchnorm transformation and ReLU as activation. We modified the last layer of the ResNet18 so it outputs the parameter of our interest (quasar SMBH mass/redshift).
• Figure 3: AGNet: A hybrid combination of a MLP and CNN.
• Figure 4: AGNet predictions for quasar SMBH redshift. Ground truth redshifts are shown by 1:1 black line with network predictions as the scatter. Trained on over $27,000$ quasars and tested on roughly $5,800$ quasars.
• Figure 5: AGNet predictions for quasar SMBH mass. Ground truth masses are shown by 1:1 black line with network predictions as the scatter. Trained on over $27,000$ quasars and tested on roughly $5,800$ quasars.