How many emission line galaxies will Euclid and Roman see?

What is it about?

The paper aims to predict the surface density and clustering bias of Hα-emitting galaxies for upcoming surveys by the Euclid and Nancy Grace Roman Space Telescope. The authors use a refined version of the GALFORM semi-analytical galaxy formation model, calibrated with deep learning emulators and Markov Chain Monte Carlo (MCMC) methods, to improve fits to observational data, including higher-redshift Hα emitter counts.

Featured Image

Why is it important?

Methodology: 1. **Model Calibration**: - Generated **3,000 GALFORM models** to train a **deep learning emulator**, enabling rapid exploration of an 11-dimensional parameter space. - Used **MCMC** to find the best-fitting model, incorporating local luminosity functions (LFs) and Hα emitter counts at higher redshifts. 2. **Emulator Design**: - Employed an ensemble of neural networks with **Leaky ReLU activation functions**, optimized for accuracy and computational efficiency. - The emulator reduced computational costs while maintaining fidelity to full GALFORM outputs. 3. **Key Parameters**: - Varied parameters related to star formation, supernova feedback, galaxy mergers, disk instabilities, and AGN feedback (e.g., \( \nu_{\text{SF}} \), \( V_{\text{SN}} \), \( \gamma_{\text{SN}} \), \( f_{\text{SMBH}} \)). Results: 1. **Hα Emitter Abundance**: - For *Euclid* (flux limit \( 2 \times 10^{-16} \, \text{erg s}^{-1} \text{cm}^{-2} \), \( 0.9 < z < 1.8 \)): Predicted **2,962–4,331 galaxies/deg²**. - For *Roman* (flux limit \( 1 \times 10^{-16} \, \text{erg s}^{-1} \text{cm}^{-2} \), \( 1.0 < z < 2.0 \)): Predicted **6,786–10,322 galaxies/deg²**. 2. Clustering Bias: - Predicted the effective linear bias of Hα emitters as a function of redshift, showing a steeper evolution at higher redshifts compared to previous studies. 3. Tensions in Calibration: - Found trade-offs between fitting local LFs and higher-redshift Hα data. Increasing weight on Hα counts improved redshift distribution fits but slightly worsened bright-end LF predictions. 4. Comparison to Observations: - The best-fitting model agreed well with recent Hα counts from Bagley et al. (2020) AND older LF data (e.g., Cole et al. 2001). Conclusions: - Demonstrated the effectiveness of **deep learning emulators** for calibrating complex galaxy formation models. - Provided updated predictions for *Euclid* and *Roman*, highlighting the importance of multi-redshift calibration. - Identified tensions between local and higher-redshift datasets, suggesting limitations in current model parameterizations.

Perspectives