Galaxy morphology as a fossil record: Katachi-PRIME infers mass, age and metallicity from three-band images

Galaxy images can serve as an information channel independent of photometry for constraining six physical quantities: stellar mass, specific star formation rate, redshift, age, dust attenuation and gas-phase metallicity. Traditional practice compresses morphology into summary statistics such as the Sérsic index, T-type or bulge fraction, or leaves morphology out of SED fitting altogether; those summaries had not strongly tightened constraints on redshift and stellar mass. On the GSWLC sample, an image model trained on gri images alone yields slightly tighter posteriors than a 5-band optical photometry model, a combined image-plus-5-band model outperforms both on population-level accuracy, and an 11-band photometry model is clearly more precise than the image model, where adding images helps little.

The method first turns each image into an embedding, then learns a one-shot mapping from data to parameter posteriors, so new galaxies need no resampling. Traditional SED fitting compares each galaxy against a library of about 1.5 million model spectra, which is slow at survey scale and beset by parameter degeneracies. A vision transformer pretrained on ImageNet and fine-tuned to predict stellar mass and star formation rate supplies 768-dimensional embeddings, which enter a neural posterior estimator built from MAF and NSF normalizing flows; training uses 185,874 galaxies and testing 46,469 mutually exclusive ones, with TARP used to check posterior convergence.

Morphology's memory of star formation is not uniform; it depends on which side of the main sequence a galaxy sits. Earlier links between morphology and star formation activity were largely qualitative, without a continuous account of how far back the record reaches. Counting a timestep as accurate when the true star formation rate falls inside the 16th-to-84th percentile posterior interval, continuous memory spans 4–7 Gyr for main-sequence galaxies, 3–5 Gyr for quiescent ones, 2–4 Gyr for green-valley ones and only 0–0.8 Gyr for starbursts; robust memory reaches 4–10 Gyr for starbursts, 7–9 Gyr for green-valley galaxies and 8–10 Gyr for quiescent ones, while the main-sequence robust memory is given as 7–9 Gyr in one passage and 8–10 Gyr in another and is not yet reconciled.

The results rest on 232,343 GSWLC galaxies with 0.02≤z≤0.16, r≤22.2 mag and GALEX plus WISE W1 coverage, using 64×64-pixel SDSS DR8 gri cutouts; training labels come from 7-band SED fitting, so the inferred quantities are SED-fit estimates rather than physical truth.

The network may lean on apparent scaling relations such as angular size and surface brightness dimming with distance to infer redshift or mass indirectly, rather than reading scale-invariant structures such as spiral arms or bars; it could also collapse into a complex color-magnitude interpolator driven by global color. Training labels lack absolute ground truth, so the model may inherit the degeneracies and systematics of traditional SED fitting, and how it performs at higher redshift, where images are less resolved, remains to be tested.