Physics-Informed Machine Learning for Urban Climate Modeling

Abstract:

Improved understanding and projection of future local urban climates under large-scale climate change are essential to sustainable urban planning and development in the long run. Urban landscape, however, is largely underrepresented in most of the state-of-the-art Earth System models (ESMs). Despite recent success on the development of local urban climate emulator using the data-driven approaches combined with dynamic simulations, less attention has been paid to embedding the physics constraints into the emulator, which may impede the predictive performance. In this work, we develop a data-driven, physics-informed machine learning approach to emulate the local urban climates. We use the large ensemble simulations by NCAR’s Community Earth System Model, which resolves the urban representation via a sufficiently physics-based urban land parameterization, as the training data. Then we impose the physics constraints (e.g., conservation of energy at the urban surface) into the machine learning-enabled model. Our framework advances the multi-model projections of the local urban climate by providing a computationally efficient and accurate emulator that can be used in conjunction with ESMs simulations.