A central barrier in improving the fundamental understanding and design of soft materials is the unavailability of transferable potential models that accurately describe inter-particle interactions in soft matter systems for a wide range of physicochemical conditions. The current fixed functional form based approaches pose limitation on the physics and chemistry that can be captured. On the other hand, fully atomistic models are limited by their short time and length scales. To address these limitations, I develop fully flexible coarse-grained deep neural network potentials that capture polymer dynamics accurately, and conduct large-scale molecular simulations for predicting macroscopic properties.