Star formation and feedback in astrophysical simulations remains a longstanding challenge when attempting to model large regions of the universe. The small scales of star formation limit the computational volume of self-consistent simulations; however, many recent works have highlighted the importance of accurate star formation and feedback to match simulations with observations. Unfortunately, the first generation of stars is often neglected in modern simulations, despite the fact that these stars generate the first metal-rich regions that fuel all subsequent star formation. This work represents the first effort to model the effect of these primordial stars without the extreme resolution requirements of prior generation simulations. We find that specialized deep convolutional neural network architectures are competent at identifying primordial star formation in under-resolved simulations, predicting star forming regions that are matched well by resolved simulations. Based on studying the {\it Phoenix Simulations}, we find that primordial star forming regions have a large, but limited influence. We generate an interpretable linear regression model to predict the size of this region based on the number, masses, and ages of stars within the primordial population.

Finally, we combine the prior works to predict primordial star formation and feedback in cosmological simulations and compare the new framework to literature-standard simulations that employ a metallicity floor. We analyze the impact of heterogeneous metal enrichment by studying the protogalaxies ($10^6 \lesssim M_v/M_\odot \lesssim 10^8$) and their stellar populations. We find that ignoring metallicity requirements for enriched star formation results in a up to $30\%$ excess in stellar mass created. Further, using a metallicity floor causes an early underproduction of stars before $z=21$ that reverses to overproduction by $z=18$, creating $\sim 20\%$ excess stellar mass and number of protogalaxies by 8.6\% by $z=14.95$. Heterogeneous metallicity conditions greatly increase the range of halo observables, e.g., stellar metallicity, stellar mass, and absolute magnitude. The increased range leads to better agreement with observations of ultra-faint dwarf galaxies when compared to metallicity-floor simulations. \starnet generates protogalaxies with $M_* \lesssim 10^3 M_\odot$, so it may additionally model low-luminosity protogalaxies more effectively than a metallicity floor criterion.