1. Introduction
Climate extremes are increasing in frequency, intensity, and complexity, creating growing environmental, economic, and societal risks in the UK and globally.
Recent events illustrate this trend, including the UK’s June 2026 heatwave, with temperatures approaching 39-40°C, following severe droughts in 2025, record winter rainfall in 2023-2024, and the unprecedented 40°C heat recorded in 2022. Globally, climate-related disasters have risen substantially over recent decades.
These accelerating extremes are rapidly pushing ecosystems into novel climate conditions that have little historical precedent. As a result, major uncertainties remain regarding how forests, grasslands, and croplands will respond, whether their capacity to absorb carbon and support food production will be maintained, and whether some ecosystems may approach critical tipping points under future climate change.
Understanding ecosystem resilience is therefore essential. Terrestrial ecosystems play a central role in climate mitigation through carbon sequestration while simultaneously providing food, biodiversity, and other ecosystem services. Improving our ability to monitor, predict, and manage ecosystem responses to climate extremes is critical for developing effective natural climate solutions and supporting long-term environmental sustainability.
Existing models were not designed for these novel conditions, and while data-driven AI methods offer new opportunities to address the problem, current methods are poorly suited to predicting into novel climate space.
Yet another practical challenge arises we often have too little data in some places and too much in others. Many key environmental processes are difficult to observe directly, while others are measured repeatedly from multiple sources and perspectives. This combination of data gaps, novel parameter space, and overwhelming complexity (i.e., the curse of dimensionality) makes it difficult for traditional models to perform reliably.
Conventional approaches in machine learning often act as “black boxes”, even when made explainable using tools like SHAP values. But these methods still do not always produce physically meaningful understanding of how ecosystems actually work; such process understanding is critical in the face of the novel climate spaces many ecosystems will face.
This PhD will develop next-generation AI methods to better understand how land ecosystems respond to climate extremes such as compound droughts and heatwaves, prolonged heavy rainfall, and changing land use pressures.
Building on the hybrid UFLUX ecosystem model (Zhu et al., 2026) developed by the supervisory team (https://github.com/songyanzhu/uflux) and efficient JAX-based machine learning tools, the project will first focus on cases where we know the general form of ecological processes, but key parameters are uncertain. For example, ecosystem respiration increases exponentially with temperature, but the exact sensitivity varies across regions and conditions.
The second and more ambitious part of the project addresses situations where we do not know the governing equations at all. For example, land-use decisions depend on a combination of environmental, policy, and socio-economic factors, but the underlying relationships are not well understood, and changing over time. Here, the project will use symbolic AI methods to automatically discover mathematical relationships directly from data. If successful, this would represent a major scientific milestone in modelling and understanding land use change dynamics (Eigenbrod et al., 2025).
2. Data and methods
The student will work with a wide range of high-resolution datasets, including:
10 years of satellite observations (via Google Earth Engine)
ground-based ecosystem flux measurements (“real-world carbon data”)
open land-use datasets (e.g. MapBiomas: https://brasil.mapbiomas.org/en/)
climate reanalysis and extreme event records
These data will be used to understand:
how vulnerable UK and global ecosystems are to climate extremes
under what conditions ecosystems may fail or recover
how land management influences resilience and carbon uptake
3. Training and supervision
This project sits at the intersection of machine learning, geography, and Earth system science. The student will be trained in:
AI for environmental systems
satellite data analysis
causal and symbolic machine learning
ecosystem modelling
The supervisory team includes:
References:
Eigenbrod, Felix, et al. "Causal machine learning methods for understanding land use and land cover change." Landscape ecology (2025).
Zhu, Zhanxing, et al. "The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects." arXiv preprint arXiv:1803.00195 (2018).
Zhu, Songyan, et al. "First global carbon dynamics from an observational and process-informed hybrid perspective: Oversimplified respiration representation likely drives divergence in terrestrial carbon sequestration across models." Agricultural and Forest Meteorology 385 (2026): 111197.