Integrating computational fluid dynamics (CFD) and machine learning to improve urban green infrastructure for heat mitigation and air quality: A systematic review

• Explored ML contributions to CFD in modeling, simulation, and strategy optimization. • Coupled CFD-ML enables real-time urban microclimate prediction and adaptive control. • Surrogate CFD-ML enables rapid UGI optimization with minimal computational overhead. • Assessed algorithm selection and applicability for integrated CFD-ML strategies. • MLP and RF dominate UGI studies for both heat mitigation and air quality modeling. Urban population growth and rapid urbanization pose significant sustainability challenges. Strategically deployed urban green infrastructure (UGI) effectively mitigates these by reducing air temperature and improving air quality. Despite advances in digital technologies, the integration of machine learning (ML) and computational fluid dynamics (CFD) to enhance UGI performance remains underexplored, with existing studies lacking comprehensive analyses of this synergy. This review systematically explores CFD-ML integration in UGI research, demonstrating how it simultaneously addresses computational efficiency and predictive accuracy in urban heat mitigation and air quality improvement. Current CFD-ML integration methods include two main frameworks: direct CFD-ML coupling and ML-based surrogate CFD modelling. Four key application areas for UGI are identified: (1) optimizing UGI design, (2) accelerating simulations, (3) improving understanding of urban microclimate physics, and (4) enhancing CFD simulation quality. The review summarizes frequently used ML algorithms, illustrating significant improvements achieved through CFD–ML integration, such as predictions up to 800 times faster with surrogate CFD-ML approaches while maintaining prediction accuracies of R² ≥ 0.90. Specifically, ML methods have accurately predicted particulate matter concentrations with errors below 10% compared to measured data, and random forest (RF) algorithms trained on ENVI-met simulations achieved cooling load prediction accuracies of up to R² = 0.98. Additionally, integrated ML frameworks significantly accelerated urban thermal analyses from over 400,000 hours to approximately one hour. Challenges remain, including coupling disparate tools, the absence of standardized or generalizable models, and data scarcity for training robust ML models. This study highlights critical research gaps, emphasizes CFD–ML integration benefits, and offers guidance on algorithm selection and integration strategies. Future research directions, including large language model (LLM)-driven CFD, hybrid physics-informed machine learning, transfer learning, and multi-fidelity modelling, are discussed as promising avenues for enhancing CFD–ML integration. Ultimately, utilizing complementary strengths of CFD and ML can produce more efficient, accurate urban climate simulations, offering substantial decision support for UGI planning in mitigating urban heat and improving air quality.