PIMALUOS: A Framework for Urban Land Use Optimization Integrating Large Language Models and Graph Neural Networks

PIMALUOS is an open-source urban land use optimization framework that adopts a three-layer Sense-Reason-Verify architecture. It integrates large language models (LLM), graph neural networks (GNN), multi-agent reinforcement learning (MARL), and physics-informed simulation technologies to achieve balanced optimization of economic, environmental, and social goals in urban land use. The framework is maintained by Parya Payami and open-sourced on GitHub.

The acceleration of global urbanization has brought complex challenges to urban land use planning. Traditional methods rely on expert experience and static models, making it difficult to address dynamic demands. The development of AI technologies (such as LLM, GNN, and reinforcement learning) has brought new possibilities to urban planning, promoting a paradigm shift from experience-driven to data-driven, and from static planning to dynamic optimization.

The core architecture of PIMALUOS draws on the cognitive science cycle:

• Sense Layer: Uses LLM to automatically parse zoning regulations and planning texts, extract constraints, and replace manual processing;
• Reason Layer: Models urban spatial relationships (adjacency, functional dependencies, etc.) via GNN, and combines MARL to achieve multi-agent collaborative decision-making;
• Verify Layer: Uses physics-informed simulation, coupled with economic, environmental, and social models to evaluate the feasibility of plans.

1. LLM-Driven Constraint Extraction: Automatically extracts planning constraints such as building height and land use compatibility from natural language texts, improving efficiency and adaptability;
2. GNN Spatial Relationship Modeling: Captures non-Euclidean relationships like adjacency and traffic connectivity of urban plots;
3. MARL Collaborative Decision-Making: Treats plots as agents, learns optimal layouts through game theory, and coordinates conflicting goals;
4. Physics Simulation Verification: Multi-dimensionally evaluates economic (output, tax), environmental (heat island effect, carbon emissions), and social (public service accessibility) indicators.

PIMALUOS is applicable to:

• Urban Planning Departments: Quickly generate alternative plans and support scenario analysis;
• Real Estate Development: Optimize land development strategies and make scientific investment decisions;
• Academic Research: Verify planning theories through large-scale experiments;
• Public Participation: Visualize plan differences to help citizens understand and participate.

The framework is built on Python and relies on a technology stack including PyTorch/Geometric (for GNN), Ray/RLLib (for distributed reinforcement learning), Transformers (for LLM interfaces), and GeoPandas (for spatial data processing). The modular design allows users to selectively enable modules, such as skipping physics simulation or connecting custom LLMs.

Limitations: Relies on high-quality data, has high computational complexity, insufficient model interpretability, and focuses mainly on static optimization; Future Directions: Introduce digital twins for dynamic optimization, develop lightweight models to support edge computing, enhance interpretability, and expand multi-modal data fusion.

PIMALUOS marks the deep application of AI in the field of urban planning, promoting the transformation of planning from an empirical art to a data science. As technology matures and data improves, intelligent planning tools are expected to become standard equipment, helping to build livable and sustainable cities. This open-source framework provides an open experimental platform for practitioners and researchers, and is worth exploring and contributing to.