The Hunt for the First Stars: How JWST and Machine Learning Search for Population III ✦
An end-to-end method has been developed to assess the detectability of the first metal-free stars (Population III) in images similar to JWST data. It combines physically motivated direct modeling based on Yggdrasil models with Bayesian inference (simulation-based inference). The analysis covers various initial mass function options, nebular properties, and Lyman-alpha transmission, accounting for noise and JADES survey filters. It is shown that unresolved analysis is severely limited by contamination from the host galaxy, while spatially resolved pixel-by-pixel model comparison significantly improves recoverability: for young massive clumps in nearly quenched galaxies at large projected distances from the center, the successful detection fraction reaches ~90%. Old and central clumps are hardly detected. Application to a candidate from the literature demonstrates spatial differentiation: the blue compact companion is better described by Population III models, while the main galaxy by standard (Population I/II).
Context
After the Big Bang, matter consisted only of hydrogen and helium; star formation began in an era called the cosmic dark ages. The first luminaries—Population III stars—lacked heavy elements and initiated nucleosynthesis, giving rise to all subsequent generations. Their detection is critical for understanding how the universe became transparent and chemically diverse. Despite their expected brightness in the ultraviolet, their short lifetimes (just a few million years) and rapid pollution by metals make these objects elusive for ordinary telescopes. Only James Webb with its infrared sensitivity can peer into the epoch when they still existed.
Methods
The researchers built a complete modeling cycle from the physics of primordial stellar spectra to realistic JWST images, using the Yggdrasil code to create spectral templates for Population III with different mass functions and nebular covering. The key innovation is the application of simulation-based Bayesian inference via neural normalizing flows, trained on millions of synthetic observations; this allows posterior probabilities for each image pixel to be estimated in a fraction of a second without directly computing the likelihood function. Morphological experiments involved numerical simulation of compact Population III clumps that ram into their host galaxies, followed by convolution with the instrumental point spread function and addition of empirical noise reproducing the conditions of the JADES deep survey. By using Hubble's law to relate redshift to distance, along with Planck's law and Stefan-Boltzmann law for radiation calculations, the authors ensured the physical fidelity of the synthetic data. A detailed scenario of nucleosynthesis in stars was first calculated in the works of Hans Bethe, Fred Hoyle and Margaret Burbidge.
Results
The analysis showed that with the classical integrated approach, the light of even a massive (10^6 M⊙) and young (1 million years) Population III clump is completely lost against the host galaxy, and standard fiducial models are always preferred. However, with pixel-by-pixel photometry, the picture changes drastically: for clumps with mass >10^6 M⊙, age <2 million years, and located at a distance >2 effective radii from the center, the model recovers the primordial nature in up to 90% of cases. Interpretation via SHAP values revealed that the dominant factors are projected distance (contribution ≈35%), redshift (≈25%), and the mass ratio of the clump to the host (≈20%); clump age and host morphology play secondary roles. Notably, on the real candidate 'Blueberry' (z=5.124), the pixel-wise Bayesian battle confirmed: the central pixel of the companion is confidently described by a Population III spectrum, while the main galaxy 'Banana' remains fiducial. This is the first time that pixel-based Bayesian diagnostics independently picked out a real object previously suspected from helium lines. In both cases, the absence of metallicity is a defining characteristic, contrasting with enriched environments containing cosmic dust.
Implications
The results significantly change the search strategy for the first generation of stars: instead of spectroscopy of faint integrated sources, one should scan the already accumulated deep fields of Webb in search of spatially resolved blue clumps on the periphery of massive galaxies at redshifts >5. Such a morphological filter sharply reduces the number of false-positive candidates caused by young low-metallicity Population II, making systematic mapping of the fleeting era of primordial nucleosynthesis realistic.
Future development
Future development of the method involves replacing simplified Sérsic profiles with realistic morphologies from cosmological hydrodynamical simulations (e.g., FLARES), thereby eliminating biases associated with prior assumptions. A breakthrough will be the implementation of amortized Bayesian model comparison via a marginal likelihood estimator, directly outputting Bayes factors and eliminating the need for post-predictive checks. Concurrently, libraries of Population III spectra will be expanded to include scenarios of partial enrichment and multi-component nebular emission.
Impact
The work will directly influence the planning of future observing programs for JWST and the Roman Space Telescope, as well as set priorities for integral field spectroscopy on 30-meter telescopes. The pixel-wise Bayesian diagnostics method will also find application in other tasks requiring the isolation of a faint component against a complex background, from searching for tidal structures to analyzing supernova remnants.
Next steps
The immediate goal is to apply the pipeline to the entire JADES archive to build a statistically significant sample of candidates, then confirm their nature with high-resolution spectroscopy. Adaptation of the pipeline to MIRI data is also needed to reach higher redshifts.
Key open problems
The search for Population III is directly tied to the mysteries of the reionization era and the origin of black holes, since primordial stars could collapse into 'seeds' of supermassive objects. The results also refine the picture of nucleosynthesis and galactic evolution, providing constraints on top-heavy mass functions.
🎯 Theoretical estimates suggest that a Population III star could be hundreds of times more massive than the Sun and explode as a hypernova, producing a flare brighter than an entire galaxy. Nevertheless, not a single confirmed representative has been found yet—their search is like chasing ghosts at the very edge of the observable universe.
Key numbers
- detection probability: up to 90% in optimal conditions
- characteristic clump mass: 10^4–10^8 M_sun
- maximum age: < 3 million years
- redshift range: 5–14
- typical pixel size: 0.03 arcsec