For spectroscopic surveys, the predictions of stellar parameters and abundances depend on a foundation of synthetic models of stellar spectra. Ideally the synthetic spectra would be indistinguishable from real spectra. The last few decades have seen significant progress towards incorporating all known stellar physics and sources of opacity in the stellar models, however there still remains a gap. All methods of decreasing this synthetic gap have so far relied on tedious by-eye comparisons of observed spectra and their synthetic counterparts. Using the latest machine learning methods, namely generative adversarial networks (GANs), our research group has been experimenting with automating the narrowing of the synthetic gap. When GANs are trained on large amounts of data, they can uncover useful information about the distribution of and patterns inherent in both the synthetic and observed domains, and the information can be used to learn about missing physics while simultaneously decreasing the synthetic gap automatically. I will present our initial results of using Cycle-StarNet (a particular type of GAN architecture) on the APOGEE Survey.