Quantifying AGN power fraction in JWST NIRCam imaging data

Pandolfo, Nadia (2025) Quantifying AGN power fraction in JWST NIRCam imaging data. Bachelor's Thesis, Astronomy.

Preview	Text NadiaPandolfoBScThesis.pdf Download (8MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (224kB)

Abstract

Disentangling host-galaxy starlight from AGN emission is key to tracing black-hole growth in the JWST COSMOS-Web survey. While Margalef-Bentabol et al. (2025) estimated AGN fractions using only the F150W band, this thesis extends the method to F115W, F277W, and F444W. Blank-sky cutouts were injected with simulated Illustris-TNG galaxies and random point-source fluxes (f_AGN = [0-0.95]), convolved with JWST PSFs. A ConvNeXt-Base model in Zoobot was trained on F277W and F444W (F115W was unstable and reserved for future work). On held-out simulations, RMSE was 0.028 and RAE 0.755 in F277W; performance in F444W was lower, in line with its broader PSF and higher thermal background. Both bands performed slightly worse than F150W in the original pipeline. Applied to real data, the network identified 3045 AGN candidates in F277W and 2110 in F444W above 5σ. Combining f_AGN estimates with Spitzer/IRAC colour cuts and a three-colour logistic classifier yields 515 median-selected, 63 MIR-selected, and 103 ML-selected AGN, with limited overlap, reflecting the complementary nature of geometric/ML and MIR selection methods. Overall, the network recovers AGN contributions with ≈0.03 absolute accuracy. Future work includes stabilising F115W training and addressing class imbalance to complete a robust four-band JWST AGN identification framework.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Wang, L. and Margalef Bentabol, B.
Degree programme:	Astronomy
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	15 Jul 2025 09:15
Last Modified:	15 Jul 2025 09:15
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36250

Actions (login required)

View Item