PhD in Climate AI: Using Geospatial Foundation Models to Refine Urban Climate Predictions

PhD in Climate AI: Using Geospatial Foundation Models to Refine Urban Climate Predictions

Quick Decision

This is a focused YES-DTN doctorate for applicants who can connect geospatial machine learning and climate science. It combines Earth-system, satellite and urban representations with generative downscaling, then tests transfer across cities using CMIP6 and Met Office 300-m data. Its strengths are a concrete research question, Leeds–Met Office collaboration and a close fit with current climate-AI work. Material uncertainties remain around CASE/placement status, international award availability and the final 2027/28 stipend. The project description lists 11 January 2027; official cohort guidance gives 13:00 GMT on 12 January 2027 and requires two forms. Plan around the earlier 11 January date until the project team confirms the applicable deadline, and confirm the funding terms before submitting.

Assessment: Strong but verify fit.

Research Problem and Doctoral Work

What Is Unsolved and Why It Matters

Global climate grids are too coarse to describe how buildings, street canyons and land cover shape local climate. Machine-learning downscaling can generate finer outputs more efficiently, but methods may fail when transferred to a city with different climate or urban form. Fine-resolution observations are also sparse, and learned patterns must be checked for physically meaningful climate effects. This gap matters because heat and rainfall risks are experienced locally while adaptation decisions need neighbourhood-scale information.

The project proposes a generative model conditioned on geospatial foundation models (GeoFMs) at three scales: Earth-system features for large-scale atmospheric context; satellite models for land cover, elevation and surface properties; and urban models for buildings, roads and socioeconomic context. These features would be integrated into architectures such as diffusion models to account for geographic context when generating high-resolution climate fields.

Work, Methods, and Potential Outputs

The plan uses global CMIP6 data and Met Office urban-climate data at 300 m. The researcher will prepare and preprocess data, prototype and train a model, integrate GeoFM representations, validate performance and test transfer across cities with different climates and urban features. The later phase aims to produce projections at 300 m or finer for selected cities under scenarios, then complete thesis and papers in the final six months of the 3.5-year timeline. Open-source methods and models are intended outputs. London, Paris and Singapore are examples, not a confirmed city set.

The central scientific test is whether spatial context improves predictions in places excluded from training. Useful checks include held-out-city evaluation, climate-shift tests, representation ablations, physical consistency, uncertainty calibration and comparisons with established downscaling baselines. These are analyst-recommended tests, not separate advertised deliverables.

The Met Office AI4Climate programme develops regional emulation, evaluation datasets and urban prediction pipelines. Existing UK climate projections include a 2.2-km local product, so the project’s 300-m scale is substantially finer. The Met Office cautions that finer resolution adds detail but does not necessarily increase confidence. The scientific value depends on validating the added detail against data and physical constraints.

School, Programme and Field Context

The University of Leeds is the degree-awarding institution and its School of Geography is the academic home. The Met Office is a research partner, not a second degree-awarding institution; no joint degree is advertised. Leeds reports Geography at #26 globally in QS Subject 2026 and Geography and Environmental Science at #14 in the UK in the Complete University Guide 2027. In REF 2021, 93% of the School’s Geography research submission was world-leading or internationally excellent, and the research environment was rated conducive to such work. These are school-level indicators, not a project rank.

The School describes a typical full-time PhD as 3–3.5 years, consistent with the project’s 3.5-year plan. Leeds provides researcher development in research integrity, data management, professional skills and careers; a training plan is due within three months of starting. Exact registration and progression requirements remain offer-specific.

Lead supervisor Dr Weiming Huang is a Lecturer in Urban Data Science and member of Leeds’s Institute for Spatial Data Science. His work includes geospatial foundation models. The listed team also includes Ben Booth and Lewis Blunn at the Met Office and John Marsh at Leeds. The combination fits the project's climate, ML and urban-geography scope. Applicants should confirm the practical arrangement for partner visits, shared supervision, compute and data access.

Supervisor and Lab Assessment

Supervisor Research Directions and Recent Papers

Three recent works by Dr Huang and collaborators give a useful view of relevant expertise:

• CityFM (2024) learns self-supervised multimodal representations from OpenStreetMap for road, building and region tasks. It is relevant to urban representations, but not climate downscaling.
• GeoFM (2025) reviews the field and identifies spatial alignment, multimodal integration, uneven data and responsible use as open problems.
• UrbanMFM (2026) integrates urban data across scales through spatial graphs and tests tasks across Singapore, New York and Beijing. It informs multiscale/cross-city representation learning, not climate transfer.

This work supports methodological fit but does not show that the proposed features already improve rainfall or temperature projections.

Lab, Resources, and Collaboration Environment

A Leeds Institute for Data Analytics project involving Huang tests graph-based representations across Leeds, London and national mobility data. Its findings vary by task, a reminder that pretrained features are not automatically informative for every target. The project says computing infrastructure will be sufficient but gives no GPU allocation, data-access procedure or software-support plan. Ask for details before accepting.

Research Direction and Field Outlook

Climate emulation and geospatial foundation models are advancing rapidly, while integrating them at city scales remains open research. ClimaX adapts a pretrained weather-and-climate model to projection and downscaling tasks; Aurora demonstrates Earth-system pretraining over multiple forecast domains; AlphaEarth supplies global satellite embeddings; and CityFM/UrbanMFM provide urban representations. These models are potential inputs or precedents, not validated 300-m urban climate predictors.

Reviews by Rampal et al. and Kendon et al. frame ML emulation as a route to faster local climate information while emphasizing scientific credibility and evaluation gaps. For this project, the key questions are whether GeoFM context adds signal beyond climate fields and conventional covariates, whether gains persist in unseen cities and future climates, and whether extremes, uncertainty and physical relationships are preserved.

Career relevance is plausible but not a guarantee. A current Met Office scientific-software vacancy in data-driven climate models lists £35,145–£37,701 and asks for ML model development, evaluation, uncertainty and physical realism. The wider current vacancies include climate-monitoring and data-engineering roles. The doctorate's coding, climate evaluation and geospatial methods could transfer to climate services, scientific software, environmental analytics and adaptation work. No project-specific graduate-outcome data was found.

Funding, Application, and Living Cost

YES-DTN says the standard 3.5-year award covers tuition, a UKRI-rate stipend and research/training/conference support. The term is 3.75 years when a three-month industry placement is undertaken. About seven of 25–26 awards may go to international applicants, with tuition covered irrespective of home/international fee difference. Relocation, visa and health-surcharge costs are excluded. The project description states that the student will be supported by a Met Office CASE award, work with its AI downscaling group and visit it regularly. The YES-DTN project profile instead describes CASE as possible and marks “Collaborative Project: No.” These sources conflict: confirm the formal classification, visits and any placement-linked extension before relying on them.

UKRI’s current minimum full-time stipend is £21,805 for 2026/27; the final 2027/28 amount is not posted. Leeds’s 2026/27 guide averages £198/week for self-catered University accommodation with bills included, plus £36 groceries and £59 personal costs: about £293/week, or £15,236 over 52 weeks. Average private rent (£210) plus groceries, non-included bills (£17) and personal spending yields about £322/week or £16,744/year. Against the current stipend benchmark, these leave about £6,569 and £5,061 respectively before irregular costs. These are guide-based student scenarios, not personal budgets or the future award amount. The stipend is maintenance support rather than salary.

The detailed YES-DTN application guide gives 12 November 2026 at 10:00 GMT as the opening, while another YES-DTN page gives 2 November 2026; confirm which opening date applies. The network deadline is 12 January 2027 at 13:00 GMT. Applicants must complete the Leeds online application and a YES-DTN combined CV/structured-statement form. The second requires the student ID issued after the first, so the programme recommends submitting the Leeds form at least two days early. Use project code YES-27-G-09 and verify it carefully. The project description's 11 January date conflicts with official instructions.

Applicant Fit and Preparation

The YES-DTN FAQ gives a relevant bachelor’s degree at 2:1 or above as a general baseline. It says first-class performance is usually needed to be competitive; a master’s is helpful but not mandatory. Listed relevant backgrounds include computer science, Earth science, environmental science and geography. The public project specifies no mandatory programming language or previous climate-modelling experience.

Applicants should demonstrate their ability to bridge computation and climate science. Computing applicants can show geospatial data work, good spatial train/test separation and interest in physical evaluation. Climate or geography applicants should show quantitative coding and ML experiments. Remote-sensing applicants should explain how features relate to climate without assuming embeddings are causal. Reproducing a small baseline and preparing a cross-city validation example are useful preparation, not added entry conditions.

Next Actions and Questions to Verify

Before Applying

• Confirm official deadline, project code, required documents and sequence of both application forms.
• Ask whether this is formally CASE/collaborative, whether the three-month placement is included or optional, and whether it adds funded time.
• Confirm international eligibility, tuition treatment, visa/health-surcharge exclusions and relocation costs.
• Ask which academic, English-language, reference and interview requirements apply to this call.

At Interview

• Which variables, cities, scenarios, baselines and spatial hold-out tests will define success?
• How will physical realism, extremes and uncertainty be tested where fine-resolution observations are sparse?
• What compute, Met Office data access, partner visits and open-source/publication terms are planned?

Before Accepting an Offer

• Confirm the written stipend, duration, fee coverage and start/end dates for 2027/28.
• Confirm Leeds registration, progression, training and the division of partner supervision.
• Clarify dataset rights, release of code/models and any partner review restrictions.

Evidence and Sources

Project and recruitment

• Official YES-DTN project profile
• YES-DTN application instructions
• YES-DTN eligibility and funding FAQ

School, supervision and training

• Leeds Geography rankings and REF profile
• Dr Weiming Huang’s Leeds profile
• School of Geography research degrees
• Leeds postgraduate researcher training
• School of Geography research environment
• University of Leeds country and contact details
• LIDA mobility foundation-model project

Scientific context and publications

• Met Office AI4Climate
• UKCP resolution and uncertainty FAQs
• Met Office urban climate impacts
• CityFM paper
• UrbanMFM paper
• GeoFM review
• Regional climate downscaling review
• Kendon et al. climate-emulator review
• ClimaX paper
• Aurora paper
• AlphaEarth paper

Funding, costs and career context

• UKRI stipend guidance
• Leeds living-cost estimates
• Met Office data-driven climate-model vacancy
• Met Office current careers and vacancies