Answers are grounded in this paper and cite their sources.
Machine learning improves power spectrum accuracy in galaxy spectra by correcting interloper impacts.
Modern slitless spectroscopic surveys such as Euclid and Roman collect vast numbers of galaxy spectra, but their low signal-to-noise ratios often lead to incorrect redshift assignments when relying on a single emission line. This misidentification, commonly caused by noise spikes or nontarget emission-line contamination, results in redshift interlopers . We present a machine learning method to mitigate interloper effects at the level of cosmological summary statistics, focusing on the power spectrum monopole as a test case. Using halo catalogs from the Quijote simulations as galaxy proxies, we simulate interlopers by displacing a fraction of halos by redshift offsets between target and interloper populations. This yields contaminated catalogs spanning a wide cosmological parameter space. We train a neural network on the power spectrum monopole, alone or combined with the bispectrum monopole, from contaminated mocks to estimate the interloper fraction and reconstruct the cleaned power spectrum. We assess performance in two scenarios: one with fixed cosmology and one where cosmological parameters vary. In the fixed case, the method recovers the interloper fraction and corrects the power spectrum to 1% accuracy. When cosmology varies, performance degrades, but including bispectrum information improves results, reducing interloper fraction errors and correcting with 2% accuracy. We also analyze how the size of the training set and the correlation between target and interloper populations affect performance. When these samples are uncorrelated, bispectrum information is more beneficial; when strongly correlated, tighter cosmological priors offer greater improvements. Our results demonstrate the potential of machine learning to enhance cosmological inference in contaminated spectroscopic data sets.
No takes yet. Share an insight, caveat, or question.
Cagliari et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: