Advances in deep learning have made constructing, training and deploying deep neural networks more accessible than ever before. Due to their flexibility and predictive accuracy, neural networks have ushered in a new wave of data-driven and data-free modeling for physical phenomena. With several key research breakthroughs in the deep learning field, modern deep learning architectures are now more accurate and generalizable facilitating improved physics-informed models. This dissertation explores the use of several different deep learning approaches for learning physical dynamics including Bayesian neural networks, generative models, physics-constrained learning and self-attention. By leveraging these recent deep neural network advancements and probabilistic frameworks, powerful deep learning surrogates of physical systems can predict complex mutli-scale features.