Designing large-scale spatial infrastructure is a longstanding challenge. Infrastructure network topologies encode implicit design principles that can be inferred from data. They are multi-scale by construction, with dense local connectivity and sparse long-range links constrained by geography and cost.
In optical backbones, physical topology directly affects network performance which in turn affects communications among people, business and organisations. Designing an optimal topology is complex due to scalability and computational constraints. Conventional graph generators rarely scale beyond ~40 nodes, while manual or optimization-based approaches require months of effort and become intractable for large-scale networks. Consequently, automating the design of optical networks remains an open challenge, further compounded by limited access to real-world deployment data.
Can graph machine learning models infer network design principles from past topologies to generate scalable topologies for the future? We introduce Topology Bench, an open dataset of 105 real-world network topologies spanning diverse regions and scales. In this work, we present Topology Architect, the first application of graph machine learning for unsupervised topology generation. We evaluate the model using graph similarity metrics, achieving 95% spectral and 84% structural similarity to real networks as measured by Wasserstein distances. The model generates topologies in under one second by sampling edge probabilities from the latent space, conditioned on user-specified parameters such as node count and geographic locations. To assess latent space quality, we compare clustering in Topology Bench using graph-theoretic properties and graph embeddings from Topology Architect, finding the latter 20 times more effective.
This research makes three key contributions: (i) provides an open-access dataset of real-world network topologies, (ii) captures the graph-theoretic properties underlying topology design, and (iii) automates topology generation. This further enables the synthesis of realistic topologies in regions with sparse or proprietary data. While developed for optical networks, this framework is transferable to large-scale infrastructure systems, including transportation, urban planning, and quantum communication.
