A Physics-Guided Deep Learning Architecture for Multi-Task Modeling of Urban Climate

Understanding climate-driven risks in urban areas necessitates improved representation of urban climate processes in Earth system models (ESMs). Machine learning (ML) has emerged as a promising tool to enhance this representation by capturing complex, nonlinear relationships in urban surfaces-atmospheric processes. Although prior ML methods achieve considerable accuracy in emulating urban climate processes for ESMs, they may compromise by focusing on a single variable and not necessarily obeying the physical laws. In contrast, physics-guided ML integrated physical elements offers promise in overcoming the above limitations, while also showcasing potential for enhanced generalization. In this work, we present a multi-task physics-guided Transformer, named UCformer, to emulate nonlinear interactions between urban surfaces and atmospheric forcing, and to generalize urban climate dynamics over time. The architecture of UCformer incorporates physical and climatic knowledge, enabling it to achieve superior performance in urban climate multi-task estimations compared to baseline models. The investigation of urban surface parameters suggests that the thermal parameters have the most integrated impact on UCformer emulation skill, with radiation parameters ranking second. Additionally, an ablation study further examines the performance gains from specific components of the UCformer architecture, underscoring the significant potential of integrating physical and climatic knowledge to enhance urban climate modeling.