Network topology synthesis seeks methods to generate large numbers of example network topologies primarily for use in simulation. It is a topic that has received much attention over the years, underlying which is a conflict between randomness and design. Random graphs are appealing because they are simple and avoid the messy details that plague real networks. However real networks are messy, because network operators design their networks in the context of complex technological constraints, costs, and goals. When random models have been used they often produce patently unrealistic networks that only match a few artificial connectivity statistics of real networks: the features that make the network useful and interesting are ignored. At best a network divorced from context is a purely mathematical object with no meaning or utility. At worst it can be completely misleading. However, design alone cannot generate an ensemble of networks with the variability needed in simulation. We need to balance design and randomness in a way that generates reasonable networks with given characteristics and predictable variability. This paper presents such a method, Combined Optimization and Layered Design (COLD), incorporating randomness and design principles to create ensembles of PoP-level synthetic networks.