Adiba Amira Siddiqa

Adiba Amira Siddiqa

Biography

Adiba Amira Siddiqa is a rising junior at Bryn Mawr College studying Astrophysics and
Applied Mathematics. Growing up as a first-generation student in Dhaka, Bangladesh, she
developed an early fascination with the Universe that inspired her to represent her country at
the International Olympiad on Astronomy & Astrophysics. Her research spans astrophysics and
machine learning, from using neural networks to identify high-redshift dark star candidates in
JWST data to modeling gravitational waves probing LISA’s sensitivity to exoplanets around
neutron-star binaries. At the Harvard-Smithsonian CfA, she built generative models recovering
galaxies’ spectroscopic properties from images and traced predictions to galaxy evolution and
anomaly detection, yielding a first-author NeurIPS workshop paper. This summer she applies deep
learning to study quasars. Beyond research, Adiba advocates STEM inclusivity on Bryn Mawr’s
Girls Who Code E-board. She finds passion in teaching, after years mentoring students, and aims to
earn an astrophysics Ph.D. to become a professor.

Interpretable Deep Learning for Black Hole Mass and Bolometric Luminosity Inference
from SDSS Quasar Spectra
Adiba Amira Siddiqa1,2, Pablo Mercader-Perez3,4, Jeroen Audenaert3 and
Anna-Christina Eilers3
1Department of Physics, Bryn Mawr College
2Department of Physics and Astronomy, Haverford College
3Kavli Institute for Astrophysics & Space Research, Massachusetts Institute of Technology
4Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology

Quasars, among the most luminous objects known, are observable across cosmic history, making
them powerful probes of black hole growth and cosmology. Their spectra encode both black hole
mass (MBH) and bolometric luminosity (LBOL). Yet, the most reliable MBH measurements come from
reverberation mapping, requiring years-long monitoring. Single-epoch scaling relations offer an
alternative but rely on single emission lines, leaving bulk of the spectrum unused. Deep learning can
exploit this information, yet it remains unclear which spectral regions these models use and
whether what they learn is physically meaningful. Here, we pretrain a conformer masked autoencoder
on 500,000 SDSS spectra; its 32-dimensional representation recovers reverberation-mapped MBH, in
cross-validation, far more reliably than catalog estimates (R² = 0.48 versus 0.26; scatter 0.47
versus 0.57 dex), outperforming AION, a far larger foundation model. Using Integrated Gradients and
traversals along the latent mass direction, we map not only which wavelengths drive each prediction
but how they change with mass: line profiles including CIV, MgII, OIII, and Hβ broaden, while
redshift-dependent features remain unchanged. Our results provide an interpretable framework for
how spectra encode physical properties, which we are extending to LBOL and generative modeling, a
step toward potentially using quasars as standard cand