PhD Final Defense ⎯ Yifan Cheng

Representing and Resolving Urban Processes at Scale

Advisor: Prof. Lei Zhao

Abstract

Cities are high-impact testbed for climate related risks with concentrated population and infrastructure. Urban surfaces alter the land-atmosphere coupling through distinct morphology-induced turbulent exchange, moisture availability, heat storage, and anthropogenic heat release. Representing urban surfaces and resolving urban processes accurately across scales are essential for analyzing urban weather, climate, air quality and extreme risks.

This dissertation builds an integrated framework to represent cities and understand their effects at scale. I achieve this goal by developing a global 1km spatially-continuous urban surface property dataset U-Surf, and by integrating U-Surf into the Community Earth System Model (CESM) to reveal the benefits brought by enhanced urban surface representation, and by developing a multi-fidelity urban surface emulator MUSE to transfer urban climate information across scales, locations and periods.

First, I leverage the latest advances in remote sensing, cloud computing and machine learning to develop U-Surf dataset based on CESM urban parameterization scheme. U-Surf, for the first time, provides the comprehensive and internally-consistent radiative, morphological and thermal properties of urban surfaces for Earth system models (ESMs). Compared to previously coarse-grained urban canopy parameters (UCPs), U-Surf demonstrates more physically-reasonably distribution ranges, preserves intra- and inter-city spatial heterogeneity even when aggregated to coarse resolutions. This dataset development work sets up the foundation for this dissertation, which marks the paradigm shift of how cities are represented in models across scales.

Next, I integrate U-Surf into CESM, which enables unprecedented global-scale urban-resolving capability in an ESM. Through comprehensive validation against global weather stations, satellite measurements, flux towers and reanalysis dataset, the land-only simulations demonstrate the benefits brought by enhanced urban surface representation in better simulating urban surface meteorology and energy fluxes. The coupled simulation further reveals the urban modulation effect on boundary layer dynamics and atmospheric processes, which in turns controls the transport, dispersion and chemical reaction of urban air pollutants. These model development works highlight the importance of scale-aware urban representation and provide broad insights for future kilometer-scale Earth system modeling.

Finally, combining U-Surf, high-resolution urban-resolving CESM simulations and urban observations, I build a multi-fidelity urban surface emulator MUSE. MUSE is pretrained on long-

term global 0.9×1.25° simulations, fine-tuned on short-term 1/32° simulations over CONUS and Europe, and adapted with urban observations. MUSE jointly predicts urban 2m air temperature, specific humidity and surface energy fluxes. The physics-informed transfer learning (PITL) framework and explicit physical constraints allow MUSE to retain the physical consistency learned from ESMs while gaining accuracy from observations. This machine learning work provides a transferrable and practical tool for urban climate studies that require kilometer scale resolution and large ensembles.

Through the comprehensive work in dataset development, Earth system modeling and machine learning, this dissertation makes urban climate modeling possible both across scales and at scale. U-Surf provide a consistent urban representation, CESM gains the urban-resolving capacity, and MUSE offers global high-resolution urban information and transferrable multi-fidelity framework. Together they lay the foundation for building scalable tools to inform decisions on heat risks, extremes, air quality and public health in cities worldwide.