As Simple as Possible but No Simpler: Optimizing the Performance of Neural Net Emulators for Galaxy SED Fitting

Artificial neural network emulators have been demonstrated to be a very computationally efficient method to rapidly generate galaxy spectral energy distributions (SEDs), for parameter inference or otherwise. Using a highly flexible and fast mathematical structure, they can learn the nontrivial relationship between input galaxy parameters and output observables. However, they do so imperfectly, and small errors in flux prediction can yield large differences in recovered parameters. In this work, we investigate the relationship between an emulator's execution time, uncertainties, correlated errors, and ability to recover accurate posteriors. We show that emulators can recover consistent results to traditional fits, with precision of $25\!-\!40\%$ in posterior medians for stellar mass, stellar metallicity, star formation rate, and stellar age. We find that emulation uncertainties scale with an emulator's width $N$ as $\propto N^{-1}$ while execution time scales as $\propto N^2$, resulting in an inherent tradeoff between execution time and emulation uncertainties. We also find that emulators with uncertainties smaller than observational uncertaities are able to recover accurate posteriors for most parameters without a significant increase in catastrophic outliers. Furthermore, we demonstrate that small architectures can produce flux residuals that have significant correlations, which can create dangerous systematic errors in colors. Finally, we show that the distributions chosen for generating training sets can have a large effect on emulators' ability to accurately fit rare objects. Selecting the optimal architecture and training set for an emulator will minimize the computational requirements for fitting near-future large-scale galaxy surveys.