Graph Neural Networks for Urban Planning: My Journey with PyTorch Geometric

romulomagalhaes

Hello fellow deep learning enthusiasts,

I’m currently diving into the fascinating world of Graph Neural Networks (GNNs) using PyTorch Geometric, and I’d love to share a unique application I’m exploring: urban subdivision planning.

Urban planning is inherently graph-structured—streets, lots, utilities, and zoning regulations form complex networks of spatial and relational data. By modeling these elements as nodes and edges, GNNs offer a powerful way to simulate, optimize, and even generate layouts that respect both engineering constraints and social dynamics.

My current focus is on:

🏙️ Representing urban lots and infrastructure as heterogeneous graphs

🧮 Using message passing to infer optimal land use and connectivity

🧠 Training models to predict development feasibility and environmental impact

PyTorch Geometric has been a game-changer for prototyping these ideas. Its modularity and integration with PyTorch’s autograd system make it ideal for experimenting with custom architectures and loss functions tailored to urban design goals.

I’d love to hear from others working on GNNs in non-traditional domains. Have you applied GNNs to spatial, civic, or infrastructure problems? Any tips for scaling models or integrating GIS data?

Let’s connect and push the boundaries of what deep learning can do for our cities.

Cheers, Romulo