Free energy is the driving force behind countless processes ranging from the biological to the industrial. Large differences in free energy drive processes forward, while large barriers impede transitions. Accurate determination of these differences and barriers allow researchers to calculate key properties. We begin with such an application. Using free energy methods in molecular dynamics, we characterize a block copolymer that forms micelles via crystallization-driven self-assembly. Through a range of free energy calculations where we determine the relative stability of micelles as a function of size, we calculate the equilibrium size, shape, and stability of these micelles. We then turn our attention to the methodology that powers this kind of analysis: free energy calculations in molecular dynamics. Given that free energy is often the quantity of interest in a system studied via molecular dynamics, the length of time these methods take to estimate the free energy strongly influences the computational cost of the studies. We present two methods that leverage self-regularizing neural networks to very rapidly estimate underlying free energy during a molecular dynamics simulation. The first method builds upon an already successful method, Adaptive Biasing Force (ABF), by better handling the error inherent in its estimates and by providing exploratory bias in unvisited regions. The second method further builds on the first by incorporating the frequency of visits in phase space, in addition to the forces, to the final estimate of the free energy for an even faster, more robust estimate. Finally, we seek to expand the reach of these methods by introducing an easy-to-use, powerful and scalable framework for applying these methods to first principles molecular dynamics, and a hierarchical transfer methodology to rapidly converge such calculations.