The next generation of ultra-bright photoemission sources may offer opportunities to enhance our understanding of fundamental spatiotemporal scales. However, modeling photoemission and laser shaping systems precisely and efficiently is difficult due to the numerous interdependent linear and nonlinear processes involved and significant variations in modeling frameworks. Here, we present a new machine learning-based framework for photoemission laser systems and dynamic laser shaping. To showcase the effectiveness of our approach in system optimization, reverse engineering, and design. Our framework is designed to facilitate precise adaptive temporal shaping, including the generation of longitudinally flat-top or periodically-modulated pulses, through integration with four-wave mixing.