From a single roof to the climate of cities

Phase one · 2009–2014Starting on the roof

In 2009, as a new doctoral student at Tsinghua University supervised by Professor Guangheng NiiD, I began measuring what a green roof actually does: not in the language of ecosystem services, but in millimetres of retained rainwater and degrees of dampened temperature swing. The experiments were small, concrete and slow. Much of my doctoral time went into that fieldwork, years before any of it could be published. A hundred square metres of sedum on a Beijing rooftop, instrumented through a monsoon season, held back around seventy per cent of the rain that fell on it and cut the daily surface-temperature swing by nearly half.

Measurement alone was never the point. In Elie Bou-ZeidiD's group at Princeton, working with Zhi-Hua WangiD, I built a model of the roof from first principles, tracking heat and water through every layer of growing medium, and used it to answer design questions a field trial could not settle: how thick should the soil be, and is irrigation worth it? The answer came with a price attached. An irrigated roof saved roughly eighteen dollars of air-conditioning over a summer. Following the physics all the way to the decision is an instinct that has stayed with me since that first instrument.

A roof is one surface. Hysteresis applies to any surface at all.

Phase two · 2014–2017Reading the city as a surface

That was the step from roof specialist to surface-energy theorist. In 2013, working with ZhihuaiD, who inspired the idea and guided me through it, I showed that the messy, empirically fitted loop between the sun's energy and the heat flowing into the ground collapses to a single physically meaningful parameter. It had been described for years by three fitted numbers that explained nothing. The result looks small on the page, and a good deal of what I have done since rests on it.

With stored heat compressed into one quantity, the next question was whether the same reasoning explained how a whole city behaves differently from the countryside. Using the 325-metre meteorological tower in Beijing and a rural reference site, in two studies led by Dan LiiD, we worked out why heat waves hit cities harder: the urban surface banks heat in concrete and brick where the countryside sheds it as evaporation, and when the wind drops the city can no longer ventilate that stored heat at night.

This phase opened the thread that runs through the rest of my career: anthropogenic heat, the warmth cities generate from buildings, traffic and bodies. We mapped it across a campus, showed it was large enough to change where rain fell over Beijing, and ran into a puzzle. Satellites appeared to see less human heat in the densest districts, exactly where there should be most. The answer, years later, was storage again: in a forest of high-rises, much of that heat is absorbed into the building fabric before it reaches air a satellite can sense.

Phase three · 2017–2022Building the instruments

By 2017 I had a set of validated physical ideas, and they lived in bespoke scripts on my own hard drive, which is where most such ideas stay. The next phase, much of it at the University of Reading alongside Sue GrimmondiD's group, was translation: turning private physics into something other people could run.

The first piece was storage heat itself. AnOHM, an analytical form of the hysteresis model, derives the storage-heat coefficients from a surface's physical properties instead of fitting them site by site, and with it I showed how heat waves change the heat a city stores, and that reflective materials and wetter soils can offset the increase.

The clearest case of translation is SuPy, my 2019 reworking of the SUEWS land-surface model. SUEWS was powerful but locked inside ageing Fortran; I wrapped it in Python so a city modeller could install it with one command. A paper like that has no headline number. Its contribution is reach, and it taught me how much leverage sits in the tool rather than the finding.

Once SUEWS could be driven from Python, it could be connected to other things. I coupled it to the WRF weather model, so a city's surface can answer back to the atmosphere above it, and tested the coupled system against flux and mixed-layer observations in London and Swindon. We derived its parameters for vegetated land at 38 FLUXNET sites, so the countryside around a city can be modelled with the same scheme, and combined it with satellite surface temperatures to map where cities store heat. It also reached building design: urban weather years generated with SUEWS give building-energy simulations the climate a building actually sits in, rather than open-terrain weather.

Around the model grew a community. SUEWS runs inside UMEP, the open toolbox that puts city climate tools in front of planners, and it took part in Urban-PLUMBER, the first major comparison of urban land-surface models in over a decade. Along the way I became one of its lead developers and its day-to-day maintainer.

Phase four · 2022–Going global, and getting specific

A city-scale model is only as good as its description of the city, and at global scale no good description of the world's buildings existed. The SHAFTS package we developed, and then the GLAMOUR dataset built on it, taught neural networks to read building heights and footprints off freely available Sentinel imagery, producing maps across more than thirteen thousand urban areas worldwide.

The period also turned towards consequence. With epidemiologists and economists, our work put a monetary figure on the urban heat island's toll across eighty-five European cities. The science also began to leave the page: our SUEWS modelling now feeds Arup's UHeat tool for mapping urban heat across global megacities, and I joined New York City's climate-budgeting advisory group, which puts heat projections into how the city spends.

One thread of recent years is general theory. In 2024, in Science Advances, a study led by DaniD made a claim about cities as a class rather than any one city: urban heat is more persistent than rural heat, and that persistence scales with the thermal inertia of the materials cities are built from. It is the 2013 hysteresis result at city scale, a statement about how long the urban surface remembers. Much of this line of work has been Dan’s.

The other thread is supervision. My name now sits more often at the end of the author list, the position of the supervisor who frames the question and guides an early-career lead. The toolkit we spent a decade building is aimed at specific decisions in specific places: compound heat risk street by street in a London borough, self-generated heat in Colombo, flood risk reconstructed for cities whose sewer maps are incomplete.

Now · 2025–The AI turn

The change I care most about now is in how the work itself gets done. The instinct that produced SuPy, that the tool matters as much as the result, is pulling me towards AI as an instrument for research rather than as a subject to model.

My main experiment is a system that turns scientific papers into claim-level knowledge graphs: verifiable statements an AI can reason over, rather than prose a person has to read end to end. It is early and unreleased. What interests me is the direction, research infrastructure in which the literature becomes computable.

It is already changing my own practice. A recent study with Chenghao WangiD, an AI-assisted systematic review of urban climate research, used LLM agents to make a review the size of a small monograph tractable, with the AI methods disclosed up front; its corpus is open to search in the IAUC Urban Climate Literature Atlas. Our review of anthropogenic heat in urban climate systems, led by DaniD, was built the same way, and what we learned from that workflow is written up as a companion piece. Closer to home, Claude Code has greatly helped me maintain SUEWS. That is what I think AI for science should look like: honest leverage on real problems, with the method stated plainly.