arXiv subjects

Daniel Grün

Publications and source records attributed to Daniel Grün.

3 recordsLinked to original sources

arXiv2026

Machine Learning Techniques for Astrophysics and Cosmology: Photometric Redshifts

The cosmological redshift of a galaxy's light is inferable from its observable properties in images. Because imaging is much easier to acquire than spectroscopic observations that would allow the identification of distinct line features, this motivates the technique of photometric redshift estimation (photo-$z$). Photo-$z$ has been an early and sustained driver for the utilization of artificial intelligence (AI) in astrophysics, and conversely AI methods are underlying most of the recent advances in photo-$z$. Here we review the diversity of AI methods applied to the photo-$z$ problem over the years in a discriminative way, that is, to regress redshift from photometric observables. We argue that, besides optimization suiting specific applications, this approach has effectively converged. It is limited not by the AI methodology but by the size and substantial systematic uncertainties and selection effects in spectroscopic training samples. In order to progress, either an unobtainable quantity and quality of training data or a more principled approach in using it is required. We thus outline ongoing research of integrating AI in a Bayesian modeling of galaxy data. This comes in the form of generative models for representing the distribution of intrinsic properties and outcomes of telescope observations of the galaxy population.

astro-ph.IM↗

arXiv2025

GalSBI-SPS: a stellar population synthesis-based galaxy population model for cosmology and galaxy evolution applications

Next generation photometric and spectroscopic surveys will enable unprecedented tests of the concordance cosmological model and of galaxy formation and evolution. Fully exploiting their potential requires a precise understanding of the selection effects on galaxies and biases on measurements of their properties, required, above all, for accurate estimates of redshift distributions n(z). Forward-modelling offers a powerful framework to simultaneously recover galaxy $n(z)$s and characterise the observed galaxy population. We present GalSBI-SPS, a new SPS-based galaxy population model that generates realistic galaxy catalogues, which we use to forward-model HSC data in the COSMOS field. GalSBI-SPS samples galaxy physical properties, computes magnitudes with ProSpect, and simulates HSC images in the COSMOS field with UFig. We measure photometric properties consistently in real data and simulations. We compare $n(z)$s, photometric and physical properties to observations and to GalSBI. GalSBI-SPS reproduces the observed grizy magnitude, colour, and size distributions down to i<23. Median differences in magnitudes and colours remain below 0.14 mag, with the model covering the full colour space spanned by HSC. Galaxy sizes are overestimated by 0.2 arcsec on average and some tension exists in the g-r colour, but the latter is comparable to that seen in GalSBI. $n(z)$s show a mild positive offset (0.01-0.08) in the mean. GalSBI-SPS qualitatively reproduces the stellar mass-SFR and size-stellar mass relations seen in COSMOS2020. GalSBI-SPS provides a realistic, survey-independent galaxy population description at a Stage-III depth using only literature-based parameters. Its predictive power will improve significantly when constrained against observed data using SBI, thereby providing accurate $n(z)$s satisfying the stringent requirements set by Stage IV surveys.

astro-ph.GA↗

arXiv2023

The BarYon CYCLE Project (ByCycle): Identifying and Localizing MgII Metal Absorbers with Machine Learning

The upcoming ByCycle project on the VISTA/4MOST multi-object spectrograph will offer new prospects of using a massive sample of $\sim 1$ million high spectral resolution ($R$ = 20,000) background quasars to map the circumgalactic metal content of foreground galaxies (observed at $R$ = 4000 - 7000), as traced by metal absorption. Such large surveys require specialized analysis methodologies. In the absence of early data, we instead produce synthetic 4MOST high-resolution fibre quasar spectra. To do so, we use the TNG50 cosmological magnetohydrodynamical simulation, combining photo-ionization post-processing and ray tracing, to capture MgII ($\lambda2796$, $\lambda2803$) absorbers. We then use this sample to train a Convolutional Neural Network (CNN) which searches for, and estimates the redshift of, MgII absorbers within these spectra. For a test sample of quasar spectra with uniformly distributed properties ($λ_{\rm{MgII,2796}}$, $\rm{EW}_{\rm{MgII,2796}}^{\rm{rest}} = 0.05 - 5.15$ Å, $\rm{SNR} = 3 - 50$), the algorithm has a robust classification accuracy of 98.6 per cent and a mean wavelength accuracy of 6.9 Å. For high signal-to-noise spectra ($\rm{SNR > 20}$), the algorithm robustly detects and localizes MgII absorbers down to equivalent widths of $\rm{EW}_{\rm{MgII,2796}}^{\rm{rest}} = 0.05$ Å. For the lowest SNR spectra ($\rm{SNR=3}$), the CNN reliably recovers and localizes EW$_{\rm{MgII,2796}}^{\rm{rest}}$ $\geq$ 0.75 Å\, absorbers. This is more than sufficient for subsequent Voigt profile fitting to characterize the detected MgII absorbers. We make the code publicly available through GitHub. Our work provides a proof-of-concept for future analyses of quasar spectra datasets numbering in the millions, soon to be delivered by the next generation of surveys.

astro-ph.GA↗