AGN -- host galaxy photometric decomposition using a fast, accurate and precise deep learning approach

TL;DR

The paper tackles the challenge of separating AGN light from host-galaxy emission in JWST imaging by introducing Zoobot, a deep learning regression model trained on mock JWST images generated from IllustrisTNG simulations to estimate the intrinsic AGN contribution fraction $f_{AGN}$ in the total light. It demonstrates that Zoobot achieves exceptional accuracy and precision (RMSE ~ 0.013, mean bias ~ -0.0018, outliers ~6.5%) and vastly outperforms traditional GALFIT-based 2D Sérsic+PSF decompositions in both accuracy and speed (∼2500× faster). The approach incorporates realistic PSF variability and applies to real COSMOS-Web data, revealing substantial but nuanced overlaps with X-ray and MIR AGN selections (e.g., 20%/8% overlaps for $f_{AGN}>0.2$, rising to 33%/15% for $f_{AGN}>0.1$). The work highlights the method’s potential for rapid, large-scale nuclear light decomposition in upcoming deep imaging surveys and outlines directions for multi-band and full galaxy–AGN separation in future studies.

Abstract

Identifying active galactic nuclei (AGN) is extremely important for understanding galaxy evolution and its connection with the assembly of supermassive black holes (SMBH). With the advent of deep and high angular resolution imaging surveys such as those conducted with the James Webb Space Telescope (JWST), it is now possible to identify galaxies with a central point source out to the very early Universe. In this study, we develop a fast, accurate and precise method to identify galaxies which host AGNs and recover the intrinsic AGN contribution to the observed total light ($f_{AGN}$). We trained a deep learning (DL) based method Zoobot to estimate the fractional contribution of a central point source to the total light. Our training sample comprises realistic mock JWST images of simulated galaxies from the IllustrisTNG cosmological hydrodynamical simulations. We injected different amounts of the real JWST point spread function (PSF) models to represent galaxies with different levels of $f_{AGN}$. We analyse the performance of our method and compare it with results obtained from the traditional light profile fitting tool GALFIT. We find excellent performance of our DL method in recovering the injected AGN fraction $f_{AGN}$, both in terms of precision and accuracy. The mean difference between the predicted and true injected $f_{AGN}$ is -0.002 and the overall root mean square error (RMSE) is 0.013. The relative absolute error (RAE) is 0.076 and the outlier (defined as predictions with RAE >20%) fraction is 6.5%. In comparison, using GALFIT on the same dataset, we achieve a mean difference of -0.02, RMSE of 0.12, RAE of 0.19 and outlier fraction of 19%. We applied our trained DL model to real JWST observations and found that 33% of X-ray-selected AGN and 15% of MIR-selected AGN are also identified as AGN using a cut at $f_{\rm AGN} > 0.1$.

Figures (21)

• Figure 1: Overview of the JWST/NIRCam F150W PSFs in COSMOS-Web. We stacked all available PSFs from Zhuang2024 and show the mean PSF (left), standard deviation (centre) and the coefficient of variation (right), calculated pixel by pixel. The PSFs have been rebinned to a pixel resolution of $0.03\ ^{\prime\prime}/\text{pixel}$, matching the resolution used for the synthetic image creation. The axes show the number of pixels, corresponding to 3$^{\prime\prime}$ across. The colorbar shows the value of each pixel.
• Figure 2: Example mock JWST/NIRCam F150W images with varying levels of AGN contribution. The images have been generated to mimic JWST observations, and include realistic JWST noise and background. Each row corresponds to a different galaxy with no AGN contribution in the left panel and increasing AGN contributions in the rest of the panels. We show four example galaxies with different magnitudes, from the brightest (top) to the faintest (bottom). Images are 3.84$^{\prime\prime}$ across and are displayed with an inverse arcsinh scaling.
• Figure 3: Architecture of ConvNeXt-Base network (bottom) with a four-stage feature hierarchy, which allows us to extract features on different scales. On top of each stage, we show the dimension of the feature maps, with the width and height decreasing as the network deepens while the filter size increases. The top left diagram shows the internal structure of ConvNeXt Block. The top right diagram shows the internal structure of Downsample. The LN and GELU represent a layer normalisation and a Gaussian error linear unit activation function, respectively.
• Figure 4: Distributions of the injected AGN contribution fraction (as defined in Eq. \ref{['eq:fAGN']}) in the training (blue) and test datasets (orange). The distributions are mostly uniform for both the training and test datasets.
• Figure 5: Sérsic + PSF decomposition of four example galaxies (with AGN contributon fraction varying from high to low from top to bottom) on which we performed GALFIT. Images of the original galaxy, the model (Sérsic + PSF), and the residual (original$-$model) are shown from left to right. Images are 3.84$^{\prime\prime}$ across, and are displayed with an inverse arcsinh scaling.