Edelen, A.L. JACoW Publishing Geneva, Switzerland The detailed design and optimization of accelerators has historically relied on high-fidelity simulations whose computational requirements limit their use as online tools. Recently, a growing community has begun reducing this computational burden by applying techniques from machine learning. For example, by learning from a sparse sampling of physics simulations one can develop fast-executing "surrogate models" that approximately predict accelerator performance for entirely new design parameters. Using these models can reduce compute times for multi-objective optimization studies by several orders of magnitude. In addition, surrogate models are now being applied in operational settings to enable non-invasive diagnostics and real-time optimization. This talk will cover developments in this field, applications to medium-energy electron photoinjectors, and how such surrogate models may improve our physics understanding of present and future accelerators.