Funded by the 1851 Royal Commission
Summary:
Urban air pollution remains one of the most pressing environmental and public health challenges facing cities globally. Predominantly driven by human activities such as transportation, industrial emissions, and residential heating, urban air pollutants significantly impact human health, contributing to respiratory illnesses, cardiovascular diseases, and premature mortality, as well as causing environmental degradation. Consequently, addressing urban air quality effectively is critical to achieving several United Nations Sustainable Development Goals (SDGs), including those related to health (SDG3), sustainable urban development (SDG11), affordable and clean energy (SDG7), and climate action (SDG13).
The current urban air quality management and policy-making processes are highly case-specific and focus on single pollutants or single effects. Consequently, policymakers must repeat a trial-and-error approach whenever making decisions in new locations. Optimal results are achieved through the co-creation of knowledge and policy options between scientists and policymakers. However, policymakers complain that the scientific contents are sometimes too complex and inaccessible for practical implementation.
The project aims to address this challenge, leading to the proposal of a two-phase policy-making framework, illustrated in the figure below. This framework comprises coarse policy-making and fine policy-making phases. By leveraging advances in Large Language Models (LLMs), a coarse policy-making process is implemented prior to fine policy-making, utilising either data-driven or model-informed, which are labour-intensive (e.g., deploying sensors across the city) and computationally expensive (e.g., running air quality models repetitively to explore the whole solution space).

Technical implementation:
In this project, we developed a workflow (as shown below), that provides a comprehensive, systematic, and technically sophisticated process designed explicitly for AI-assisted urban air quality policy-making. This workflow transforms complex and voluminous scientific data into actionable, quantitatively validated insights by integrating structured computational knowledge extraction, semantic querying, predictive modelling, and interactive policy assessment.

The workflow operationalises the two-phase framework through clearly defined technical steps. The initial coarse policy-making phase systematically extracts and organises domain-specific knowledge from scientific literature, technical reports, and official policy documents. Employing Large Language Models (LLMs) combined with a Retrieval-Augmented Generation (RAG) mechanism, this stage constructs a structured knowledge graph implemented in Neo4j. The graph, enriched with semantic embeddings (generated using SentenceTransformers), enables accurate retrieval of relevant air quality interventions, emission sources, associated pollutants, and documented policy outcomes. By grounding AI-generated insights directly in authoritative sources, the RAG technique ensures the recommendations’ accuracy, transparency, and credibility.
*Please refer to https://github.com/XiangX91/urban-air-quality-kg for the implementaion of the coarse policy-making.
The subsequent fine policy-making phase employs rigorous quantitative analyses through advanced simulation modelling. Specifically, the workflow leverages the SHERPA (Screening for High Emission Reduction Potential on Air) modelling framework, facilitating rapid yet precise evaluation of air quality policy impacts. Central to this phase is the detailed source-allocation analysis, quantitatively assessing how distinct emission sectors—such as transportation, industrial activities, or residential heating—uniquely affect local air quality. This source-specific analysis refines and validates the broader insights derived from the coarse phase, allowing for tailored and precise policy recommendations suited to particular urban contexts.
*Please refer to https://github.com/XiangX91/urban-air-mitigation-sim for the implementaion of the fine policy-making.
Finally, the workflow integrates qualitative insights and quantitative simulation results into an interactive decision-support environment. Policymakers can intuitively explore scenarios through natural-language queries and visual interfaces, dynamically examining policy outcomes. Moreover, user feedback gathered during policy evaluations continually refines both the knowledge graph and the simulation model parameters, thus ensuring the decision-support system’s sustained relevance, precision, and practical utility.