Physics-informed Machine Learning for decision making related to future extreme weather events

Year
2025
Primary Supervisor
Stefan Siegert <s.siegert@exeter.ac.uk>
Institution
University of Exeter (Mathematics and Statistics)
Academic Supervisors
Jemma Shipton <J.Shipton@exeter.ac.uk> (University of Exeter, Mathematics and Statistics)
Project Partners
KAUST (King Abdullah University of Science and Technology)

Background

Climate risk analysis aims to understand and predict extreme weather events, such as severe precipitation, which might become more frequent and intense due to climate change. Probabilistic machine learning (ML) models, especially those capable of producing realistic spatiotemporal output, are useful tools to estimate the risk of short-lived localised weather events based on large scale climate variables. A recent advancement in data-driven ML is the inclusion of physical principles, such as physics-informed neural networks (PINNs), to improve model accuracy and realism by constraining model output to obey known physical laws. Physics-based ML models can potentially capture the complex processes that drive severe weather, but many challenges remain in ensuring their practical applicability. Ongoing research aims to refine physics-based probabilistic ML models to support decision-making processes in risk assessment, disaster preparedness, and insurance.

PhD opportunity

Challenge 1) We have to decide on how exactly physical mechanisms should be included in a probabilistic ML model conditioned on large scale climate variables. The model can be physics-constrained by a) augmenting training data with idealised simulations; b) by hard-coding physical constraints in the ML model architecture itself; or c) by modifying the loss function to penalise deviations from physical constraints. The project will conduct a systematic review and comparison of methods with focus on skill improvements of spatiotemporal predictions of extreme precipitation events in idealised synthetic data and on real-world data sets.
Challenge 2) We need robust methods for generating spatiotemporal probabilistic predictions from neural networks and other ML methods, which is crucial for accurately assessing climate risks. This project will explore and refine approaches such as ensemble-based techniques, diffusion models, and custom loss functions. The aim is to conduct a systematic review of existing methods, followed by the development of novel solutions to improve the accuracy and reliability of probabilistic forecasts of magnitude, duration and spatial extent for extreme precipitation events. These outcomes will support decision-making processes in climate risk assessment and enhance our understanding of prediction uncertainty.
Challenge 3) Robust evaluation methods are essential for selecting the best ML model and assessing its utility in climate risk contexts. This project will focus on developing and refining evaluation methods for probabilistic predictions, incorporating metrics sensitive to the accuracy, reliability, and practical utility of predictions in real-world risk assessment scenarios. By systematically evaluating model performance across different climate risk contexts, this work will contribute insights into model selection and risk communication, enabling more informed, data-driven decision-making under climate uncertainty.

Other information

Applicant profile: Students with a strong background in mathematics, statistics or physics who are keen to apply their skills to complex, real-world challenges in climate science and risk assessment are well-suited for this PhD. Ideal applicants will have experience or a keen interest in machine learning, data science, or probabilistic modelling, particularly in spatiotemporal contexts. Strong analytical skills, proficiency in programming (e.g., Python, R), and an interest in interdisciplinary research that connects mathematical/physical sciences with decision-making are essential. Applicants should be motivated to explore new methods for improving climate risk assessment and eager to contribute to developments in applied climate risk assessment and decision science.

  • “A Review of Recent and Emerging Machine Learning Applications for Climate Variability and Weather Phenomena” https://journals.ametsoc.org/view/journals/aies/2/4/AIES-D-22-0086.1.xml
  • “Physics-informed machine learning” https://www.nature.com/articles/s42254-021-00314-5