How AI can restructure urban layouts to slash transport emissions

The AI’s learned behavior shows a clear preference for high-density, mixed-use development, increasing the spatial clustering of residential areas by 23 percent and commercial areas by 31 percent. By concentrating key land uses in walkable proximities, the model sharply reduces transportation-related emissions, which represent one of the largest contributors to urban carbon output.

A new scientific investigation reveals that deep reinforcement learning could become a transformative tool for governments attempting to curb carbon emissions through urban land-use planning. The study demonstrates that artificial intelligence (AI) can restructure land distribution patterns to significantly lower emissions while supporting more sustainable urban development.

The findings come from the study “Optimizing Urban Land-Use Through Deep Reinforcement Learning: A Case Study in Hangzhou for Reducing Carbon Emissions,” published in Land. The team developed a novel AI-based land-use optimization model trained on Points of Interest (POI), Areas of Interest (AOI), and transportation system data to identify the configurations that most effectively reduce carbon output across Hangzhou, one of China’s fastest-growing metropolitan regions.

AI outperforms conventional methods in urban emission reduction

Expanding built-up areas, increasing transport demands, and shrinking natural carbon sinks create conditions in which traditional land-use planning struggles to maintain environmental balance. The authors note that existing machine learning tools can estimate emissions but cannot handle the sequential decision-making required for citywide land-use optimization.

To bridge this gap, the researchers designed a Deep Reinforcement Learning (DRL) framework based on the Proximal Policy Optimization (PPO) algorithm, enabling an AI agent to interact dynamically with a simulated urban environment. The agent modifies land-use proportions across the city grid, receives feedback on resulting emissions, and gradually learns the most effective configurations for reducing carbon output.

The Hangzhou case study divides the city into 500-meter grid cells, with each cell assigned a land-use distribution comprising residential, commercial, industrial, recreational, and public service categories. Using the AOI, POI, and transportation data, the authors constructed a detailed carbon emission model that captures two critical sources: emissions produced by each land-use type within a grid cell, and the emissions generated by movement between cells based on urban mobility patterns.

The AI model ran over 400 training episodes, continually updating its strategies to maximize emission reductions. The results show that DRL outperformed both linear regression and genetic algorithm baselines by a significant margin, demonstrating superior accuracy, robustness, and stability in identifying optimal spatial configurations.

Across Hangzhou, the DRL model achieved up to a 15 percent reduction in total carbon emissions, a substantial improvement when scaled to citywide energy use and population density. In contrast, the linear regression baseline produced only modest emission reductions, while the genetic algorithm achieved moderate but inconsistent results. The study emphasizes that this performance gap underscores DRL’s potential as a decision-support tool in climate-focused planning.

The AI’s learned behavior shows a clear preference for high-density, mixed-use development, increasing the spatial clustering of residential areas by 23 percent and commercial areas by 31 percent. By concentrating key land uses in walkable proximities, the model sharply reduces transportation-related emissions, which represent one of the largest contributors to urban carbon output.

This shift toward compact urban forms aligns with global research showing that mixed-use development shortens commute times, reduces congestion, and increases the efficient deployment of public services. The authors argue that DRL provides empirical support for city planners who advocate for higher-density growth patterns but lack actionable computational tools to substantiate long-term environmental outcomes.

AI learns environmentally efficient spatial patterns through iterative decision-making

The strength of the DRL framework lies in its ability to learn through interaction, rather than relying on fixed statistical relationships. Unlike static regression models, the PPO-based system continuously tests hypothetical land-use changes, evaluates emission outcomes, and improves its decision-making over time. The model can account for the interconnected nature of urban systems, where a change in one cell may ripple through transportation networks and alter emissions elsewhere.

The study highlights four layers of data that make this possible:

• AOI data, which capture broader functional zones
• POI data, which reflect fine-grained human activity patterns
• Transportation system data, which indicate mobility intensity and carbon cost
• Land-use proportions, which define functional composition within each grid

By integrating these layers, the model learns both base emissions within each grid cell and path emissions between cells. The latter is crucial, as transportation accounts for a substantial share of urban carbon output.

The authors note that the model consistently encourages denser clustering of residential, dining, and commercial zones, revealing the AI’s emphasis on minimizing travel distances. This is consistent with established sustainable development principles but achieved here through autonomous learning rather than manually defined planning rules.

The model also demonstrates sensitivity to the spatial distribution of public services, balancing accessibility with energy efficiency. For example, dispersed medical or educational facilities may increase accessibility but also increase travel emissions. The PPO agent identifies configurations that reduce emissions without eliminating essential public infrastructure.

Compared with genetic algorithms, often used in urban simulation, the DRL framework displays more stable convergence behavior, avoiding random volatility and delivering consistent performance across multiple iterations. This reliability is essential for urban planners seeking predictive tools that support long-term policy development.

The study notes that while the model is highly effective, it remains computationally intensive and dependent on high-quality spatial datasets. However, as more cities adopt digital mapping and real-time mobility tracking, the potential for DRL-based planning will expand rapidly.

Limitations and future directions for real-world integration

While the study demonstrates the power of DRL for emission-oriented land-use planning, the authors acknowledge several limitations that must be addressed before such models can be deployed widely in policymaking.

The first challenge is that the simulation environment is deterministic, meaning it does not account for unpredictable real-world conditions, such as economic shocks, policy interventions, population changes, or public resistance to rezoning. Future models will need stochastic elements to simulate uncertainty and provide more resilient planning recommendations.

The study also notes the absence of socioeconomic constraints, such as housing affordability, equity considerations, community displacement risk, and political feasibility. While the DRL agent optimizes for emissions, real cities must satisfy multiple, often conflicting objectives. Multi-objective reinforcement learning could enable planners to balance environmental, social, and economic outcomes.

The generalizability challenge is another concern. The framework was trained on Hangzhou’s unique geographic, economic, and demographic conditions. Applying the model to other cities will require recalibrated emission factors, local transportation patterns, and cultural activity data.

Additionally, the study recognizes that carbon emission factors used in the model may change over time due to advances in vehicle efficiency, electrification, or renewable energy adoption. Dynamic emission coefficients will be needed for models intended to support long-term planning.

Data availability poses structural constraints as well. Many cities worldwide lack the granular POI, AOI, and mobility datasets required to build such models, especially in regions where digital infrastructure is still developing. Expanding open-access urban data ecosystems will be essential if AI-driven land-use optimization is to be applied globally.

Despite these challenges, the authors argue that DRL offers a compelling foundation for future planning technologies.