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

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 $f_{\rm AGN} > 0.1$.

💡 Research Summary

This paper presents a novel, deep‑learning (DL) approach for photometric decomposition of active galactic nuclei (AGN) and their host galaxies in high‑resolution JWST imaging. The authors aim to recover the fractional contribution of a central point source to the total galaxy light, denoted $f_{\rm AGN}$, with higher accuracy and speed than traditional two‑dimensional surface‑brightness fitting tools such as GALFIT.

Data and Simulations
The observational dataset consists of JWST/NIRCam F150W images from the COSMOS‑WEB treasury survey, reduced with the official JWST pipeline and accompanied by a set of 80 empirically derived PSF models that capture both temporal and spatial variations (median FWHM ≈ 61 mas). To build a realistic training set, the authors select galaxies from the IllustrisTNG cosmological hydrodynamical simulation, spanning a wide range of stellar masses, redshifts, morphologies (including mergers), and sizes. For each simulated galaxy they generate mock JWST images, inject the real JWST PSF at varying amplitudes, and thus create a library of images with known $f_{\rm AGN}$ values ranging from 0 to 0.9 in ten steps. The mock images preserve the noise characteristics, background level, and PSF variability of the real data, ensuring that the DL model learns to handle realistic observational conditions.

Deep‑Learning Method
The authors adopt Zoobot, a convolutional neural network originally designed for galaxy morphology classification, and repurpose it for regression of $f_{\rm AGN}$. Input images are 64 × 64 pixel cutouts; the network outputs a single continuous value. Training uses mean‑squared error loss, the Adam optimizer (learning rate 1e‑4), and early stopping based on a validation set. Data augmentation (rotations, flips, scaling) mitigates over‑fitting. The model is trained on ~100 k mock images and validated on an independent set that was not used during training.

Performance Evaluation
On the validation set the DL model achieves a mean bias of –0.002, root‑mean‑square error (RMSE) of 0.013, relative absolute error (RAE) of 0.076, and an outlier fraction (RAE > 20 %) of only 6.5 %. By contrast, GALFIT applied to the same images yields a mean bias of –0.02, RMSE = 0.12, RAE = 0.19, and an outlier fraction of 19 %. The DL method remains robust across a range of injected $f_{\rm AGN}$, redshift (z ≈ 0.5–3), signal‑to‑noise ratio, and effective radius, whereas GALFIT struggles especially for low S/N, highly disturbed morphologies, or when the PSF model is mismatched. Importantly, the DL inference time per galaxy is < 0.01 s on a modern GPU, compared with several seconds to minutes per object for GALFIT, highlighting the computational advantage for large surveys.

Application to Real JWST Data
The trained model is applied to real JWST/F150W images in the COSMOS‑WEB field. Two independent AGN samples are used for cross‑validation: (1) X‑ray‑selected AGN from deep Chandra observations, and (2) mid‑infrared (MIR) selected AGN from Spitzer/IRAC. Using a threshold $f_{\rm AGN}>0.1$, the DL method identifies 33 % of the X‑ray AGN and 15 % of the MIR AGN as having a significant point‑source component, demonstrating that image‑based decomposition can recover AGN missed by wavelength‑based selections alone. The method also shows a high success rate (> 95 %) for galaxies with irregular or merging morphologies, where GALFIT often fails to converge.

Discussion and Limitations
The authors discuss several strengths: (i) automatic handling of PSF variability without explicit PSF fitting, (ii) independence from assumed analytic surface‑brightness models (e.g., Sérsic), (iii) scalability to massive imaging datasets, and (iv) consistent performance across diverse galaxy types. Limitations include reliance on the realism of the simulated training set (e.g., dust geometry, sub‑kiloparsec clumps may differ from real galaxies), reduced accuracy for extremely high AGN fractions ($f_{\rm AGN}>0.8$) where the host becomes indistinguishable, and the current use of a single filter. Future work is proposed to incorporate multiple JWST filters, explore other cosmological simulations (EAGLE, SIMBA) for cross‑training, and embed Bayesian uncertainty estimation within the network.

Conclusions
The study demonstrates that a deep‑learning model trained on realistic JWST mock images can recover AGN light fractions with an order‑of‑magnitude improvement over traditional GALFIT fitting, while being orders of magnitude faster. This approach opens the door to automated, high‑precision AGN–host decomposition in forthcoming large‑scale JWST and next‑generation optical/IR surveys, enabling more reliable studies of SMBH–galaxy co‑evolution across cosmic time.