Topology Architect: Graph Generative Intelligence for Scalable Optical Network Topology Design

The design of an optical network's physical topology determines its performance through its graph properties. However, the intelligent and automated design of scalable optical networks remains challenging due to the computational complexity of traditional optimisation methods, deterministic graph generators, and the lack of proper training data. In this paper, we introduce Topology Architect, the first generative AI model for optical network design, capable of generating core topologies from only node counts and geographic coordinates. It is an unsupervised learning framework trained on real networks from our dataset, Topology Bench. The model achieves up to 95% graph similarity, as measured by Wasserstein distances to real networks, and generates user-defined topologies in less than a second. Topology Architect captures the data variability in the latent space 20 times better than a graph's mathematical properties alone. It implicitly integrates multi-objective design principles, scales across different sizes, and presents a novel framework for optical network topology generation.