pop-cosmos: Star formation over 12 Gyr from generative modelling of a deep infrared-selected galaxy catalogue

We study star formation over 12 Gyr using pop-cosmos, a generative model trained on 26-band photometry of 420,000 COSMOS2020 galaxies (IRAC Ch.1 $<26$). The model learns distributions over 16 SPS parameters via score-based diffusion, matching observed colours and magnitudes. We compute the star formation rate density (SFRD) to $z=3.5$ by directly integrating individual galaxy SFRs. The SFRD peaks at $z=1.3\pm0.1$, with peak value $0.08\pm0.01$ M$_{\odot}$ yr$^{-1}$ Mpc$^{-3}$. We classify star-forming (SF) and quiescent (Q) galaxies using specific SFR $<10^{-11}$ yr$^{-1}$, comparing with $NUVrJ$ colour selection. The sSFR criterion yields up to 20% smaller quiescent fractions across $0<z<3.5$, with $NUVrJ$-selected samples contaminated by galaxies with sSFR up to $10^{-9}$ yr$^{-1}$. Our sSFR-selected stellar mass function shows a negligible number density of low-mass ($<10^{9.5}$ M$_\odot$) Q galaxies at $z\sim1$, where colour-selection shows a prominent increase. Non-parametric star formation histories around the SFRD peak reveal distinct patterns: SF galaxies show gradually decreasing SFR correlations with lookback time ($r\sim1$ to $r\sim0$ over 13 Gyr), implying increasingly stochastic star formation toward early epochs. Q galaxies exhibit full correlation ($r>0.95$) during the most recent $\sim$300 Myr, then sharp decorrelation with earlier star-forming epochs, marking clear quenching transitions. Massive ($10<\log_{10}(M_*/$M$_{\odot})<11$) galaxies quench on a time-scale of $\sim1$ Gyr, with mass assembly concentrated in their first 3.5 Gyr. Finally, AGN activity (infrared luminosity) peaks as massive ($\sim10^{10.5}$ M$_\odot$) galaxies approach the transition between star-forming and quiescent states, declining sharply once quiescence is established. This provides evidence that AGN feedback operates in a critical regime during the $\sim1$ Gyr quenching transition.