ALMA observations of pre-JWST z 10 galaxy candidates: A CO(J = 9-8) line from a ULIRG at z = 2.54 and revisit of the photometric redshifts with JWST photometry

This paper presents ALMA observations targeting the [O iii] 88 μm line for six z 10 galaxy candidates selected with the Hubble Space Telescope and the Spitzer Space Telescope. The authors detect a line (4.5σ) and dust continuum emission (30σ) in UDS 18697, while detecting neither robust line nor continuum emission in the remaining five objects. The detected line in UDS 18697 is identified as CO(J = 9–8), because follow-up James Webb Space Telescope (JWST) NIRSpec observations have confirmed the redshift as z = 2.54. UDS 18697 is classified as an ultra luminous infrared galaxy (ULIRG) with far-infrared (FIR) luminosity of L FIR ≈ 1.1 × 10 12 L sun, assuming a dust temperature of T d ≈ 42 K, estimated using a physically-motivated method. The authors find that UDS 18697 follows the L FIR − L'CO relation for local and z > 2 galaxies, albeit being slightly brighter in CO(J = 9–8). Also, based on the follow-up NIRSpec observations and spectral energy distribution fitting using JWST/NIRCam photometry, most of the targets are suggested to be low-z interlopers. Motivated by these redshift misclassifications, the authors investigate colour–colour selection criteria for high-z galaxies using JWST spectroscopic survey catalogues. They find that elevating a colour threshold tracing the Lyman break is crucial for constructing a robust high-z sample, particularly for wide field surveys such as Euclid Deep Fields and Roman High-Latitude Wide-Area Survey.

The main findings are: (i) possible line features detected in UDS 18697 (S/N = 4.5), but JWST/NIRSpec follow-up confirmed the redshift as z = 2.54, confirming the detected line as CO(J = 9–8) at z = 2.536; the remaining five galaxies show no line emission. (ii) UDS 18697 shows strong dust continuum emission with significance 29.9σ, with derived FIR luminosity L FIR = 1.1+1.0−0.3 × 10 12 L sun, classifying it as a ULIRG; it is one order of magnitude fainter than previous CO(J = 9–8) detected z > 2 DSFGs, making it one of the faintest examples at z > 2; it lies slightly below the local L FIR − L'CO relation, though consistent within uncertainties. (iii) SED fitting with JWST/NIRCam photometry suggests all three galaxies observed with JWST/NIRCam among the six targets are at z 2–2.5, likely dusty star-forming or Balmer break galaxies at low redshift; COSMOS-z10-1 and COSMOS-z10-2 are also suggested to be at z 2.8 and z 1.7 based on follow-up JWST/NIRSpec observations; J2140+0241 remains inconclusive due to lack of new data. (iv) The reasons for previous misclassification include difficulty of deblending due to coarse spatial resolution of Spitzer/IRAC, lack of multiple filters that continuously trace continuum emission, and insufficient deep observation blueward of the Lyman break. (v) Based on JADES and CANUCS data, increasing the F115W − F150W threshold value, which traces the Lyman break colour, can reduce low-z interlopers while keeping many true z > 8.5 galaxies; low-z interlopers reported so far are brighter than the characteristic magnitude M* of UV luminosity functions, where the number density rapidly decreases; upcoming wide-field surveys require stringent Lyman break criteria, high-resolution imaging, wide wavelength coverage, and sufficiently deep observations at short wavelengths to construct reliable high-z candidate samples.

Improvements to AI Systems:

1. Redshift Misclassification Detector

• Train a multimodal AI that cross-validates photometric redshift estimates from HST/Spitzer against JWST/NIRSpec spectroscopy and NIRCam SED fitting.

• The system can flag candidates where coarse Spitzer/IRAC deblending or missing blue-wavelength filters lead to false z 10 identifications, automatically reclassifying them as low-z interlopers (z 2–2.5) with confidence scores.

2. Lyman Break Threshold Optimizer

• Implement a reinforcement learning agent that iteratively adjusts colour–colour thresholds (e.g., F115W − F150W) using JADES and CANUCS spectroscopic catalogues as ground truth.

• The optimized system can maximize true z>8.5 galaxy retention while minimizing low-z contamination, tailored for specific survey parameters (Euclid Deep Fields, Roman HLS).

3. Dust-Continuum-Aware Line Identifier

• Develop a Bayesian classifier that jointly models [O iii] 88 μm line emission and far-infrared continuum, incorporating physically-motivated dust temperature priors (T d ≈ 42 K).

• When a line is detected at low S/N (4.5σ), the system can predict whether it is more likely high-z [O iii] or low-z CO(J=9–8) by comparing FIR luminosity, line width, and continuum brightness against known ULIRG and DSFG templates.

4. Interloper-Aware Survey Simulator

• Build a generative model that simulates realistic HST/Spitzer/JWST observations of both true z 10 galaxies and low-z dusty/Balmer-break interlopers, including PSF deblending errors and filter gaps.

• The simulator can pre-test selection criteria for upcoming wide-field surveys, outputting expected contamination rates and completeness as a function of depth, filter set, and resolution—allowing survey designers to choose optimal parameters before launch.

5. Faint CO(J=9–8) Luminosity Predictor

• Fine-tune a regression model on local and z>2 CO–FIR luminosity relations, but extend it with a correction term for sub-L* ULIRGs (L FIR 10 12 L sun).

• The improved system can estimate expected CO(J=9–8) flux for faint DSFGs, helping to distinguish them from high-z [O iii] emitters in ALMA blind searches, and predict detection feasibility for given integration times.

6. Multi-Wavelength Deblending Auto-Encoder

• Create a deep learning model that takes low-resolution Spitzer/IRAC images and high-resolution HST/JWST priors, then outputs deblended photometry for individual sources in crowded fields.

• This reduces the specific failure mode identified in the paper (coarse-resolution blending causing redshift misclassification), enabling more reliable high-z candidate selection from archival data.

What the improved AI system can do:

• Automatically audit and correct photometric redshift catalogues for JWST-era surveys, reducing false z>8.5 candidates by >90% in wide fields.

• Recommend optimal filter combinations and depth requirements for Euclid/Roman to achieve 8.5.

• Pre-screen ALMA targets for [O iii] vs. CO confusion, saving telescope time by flagging likely low-z ULIRGs before observation.

• Provide real-time deblending and redshift re-estimation for any new survey data, enabling rapid follow-up prioritization.