Machine learning is becoming an integral part of biomolecular modeling and scientific discovery generally. Recently, the 2024 Nobel Prize in Chemistry highlighted the role that deep learning models have played in predicting protein structure from amino acid sequence, but solving static structure is just one piece of the biomolecular puzzle. Physics-based simulations play a crucial role in determining the thermodynamics, kinetics, and mechanisms of diverse molecular systems. Despite improvements in force-field accuracy and computational power, these simulations remain limited by the time and length scales they can reliably sample. We see immense opportunity for deep learning models to enhance molecular simulations and to extend their application beyond traditional spatiotemporal limitations. This dissertation addresses three missing links in computational molecular engineering and molecular dynamics simulation: i) using insights from coarse-grained simulations to interpret experimental observables, ii) accessing biologically relevant timescales from inherently short computational timescales, and iii) backmapping all-atom molecular structures from low-resolution simulation trajectories. For each of these goals, we leverage the power of deep learning as a universal function approximator to perform a diverse set of tasks ranging from optimizing dynamical embeddings to generating synthetic trajectories to denoising atomic coordinates. This work and the associated publications represent both a scientific investigation of specific biophysical processes as well as a collection of methodological developments which can be applied to a diverse set of molecular systems.