Aryana Haghjoo

Super-resolving low-resolution galaxy spectra (SR2)

Low-resolution spectra from JWST/NIRSpec's prism (R∼30–300) blend together key rest-frame optical lines — Hα, Hβ, [OIII]λ5007, [OII]λ3727 — past the point where dust attenuation, metallicity, and ionization state can be read off directly. Grating follow-up resolves them, but it's expensive and only available for a fraction of targets. I built a three-stage transformer pipeline (which I call SR2) that takes a low-resolution prism spectrum and reconstructs it at roughly 10× higher resolving power (R∼100 → R∼1000), trained on 1,187 paired prism/grating observations from the JADES program.

The pipeline first performs a coarse convolutional super-resolution pass, then infers a redshift from that intermediate reconstruction, then applies a physics-informed residual correction that uses self-attention across emission-line tokens to learn how lines relate to one another, alongside a convolutional branch that handles the continuum. On held-out test spectra it reaches noise-limited residuals across most of the spectral range and successfully deblends features — like the [OIII]λλ4959,5007 doublet — that are completely unresolved at prism resolution, with meaningful signal-to-noise improvements on all four diagnostic lines.

I've since stress-tested SR2 against seven classical deconvolution methods (Wiener, Tikhonov, total-variation, Richardson–Lucy, wavelet-sparse, and a redshift-informed matched filter) on the same 1,187-spectrum set, to find out exactly where the deep-learning advantage is real versus where it's an illusion of denoising. SR2 wins decisively on global fidelity (30% lower error than the best classical method) and on line-width recovery (the smallest FWHM bias of any method, on every line) — genuine super-resolution, not just noise suppression. But the comparison also surfaced a real limitation: SR2 is the worst method of the eight for preserving flux ratios like the Balmer decrement and [OIII]/Hβ, because it denoises Hα more aggressively than the fainter lines. That matters, because those ratios are exactly what astronomers use to measure dust attenuation and ionization state. The practical takeaway is that SR2 is the right tool for detecting and characterizing line shapes, but flux-ratio science still needs either classical reconstructions or a calibration step on top of the network's output. I'm currently extending the architecture to fix this, and separately toward a multi-modal framework that fuses photometric and spectroscopic data for cross-modal inference.

The next direction is moving SR2 beyond JWST entirely: we're working on transferring what the model learned on NIRSpec prism/grating pairs to the Nancy Grace Roman Space Telescope and Euclid, two upcoming wide-field missions that will produce enormous volumes of low-resolution slitless spectroscopy with little to no high-resolution follow-up to train against. The open question is how much of SR2's physics-informed prior — the emission-line relationships it learned from JWST — carries over to a different instrument, a different wavelength coverage, and a different survey strategy, and how to adapt the model where it doesn't.

Advised by Bahram Mobasher & Shoubaneh Hemmati — UC Riverside

Parameter estimation of the global 21cm signal

The global 21cm signal is one of the few direct observational windows into the period between the end of the cosmic dark ages and the formation of the first stars and galaxies — it's sensitive to the density and temperature of neutral hydrogen at that time, which means any deviation from what the standard cosmological model predicts could be a sign of new physics: cosmic strings, exotic particle interactions, unconventional dark matter candidates.

My thesis built a parameter estimation pipeline combining Markov Chain Monte Carlo with the Levenberg–Marquardt algorithm to recover the best-fit astrophysical parameters of the 21cm curve from theoretical models generated with Accelerated Reionization Era Simulations (ARES), then validated the approach against real data from the EDGES experiment. The literature review underpinning it covers the physical mechanisms that shape the signal, how non-standard physics would imprint on it, and the practical difficulties — foreground removal especially — that make this signal so hard to observe in the first place. The resulting best-fit parameters are meant to help constrain future theoretical models and set realistic precision targets for the next generation of experiments hunting for these non-standard effects.

Advised by Jonathan Sievers & Oscar Hernandez — McGill University

Classifying variable stars with machine learning

Applied KNN, random forest, and CNN models to classify variable stars in the OGLE dataset, working through the challenges of significant class imbalance.

Advised by Sadegh Raeisi & Sohrab Rahvar — Sharif University of Technology