2024-B30 Machine Learning for Long Range Forecasts

Project highlights

Opportunity to drive the science of seasonal to decadal forecasting forward in collaboration with world-leading forecast producing partner, the UK Met Office, our project CASE partner.
Learn data science skills, ML-approaches, and the development and use of cutting-edge long-rage forecasts.
The student will have a systematic engagement with the Met Office by visiting and working at the Met Office for certain times, thus profiting form the excellent research environment at the Met Office as well.

Overview

Seasonal and decadal ensemble forecasts can provide long range advance warnings of climate variations and can estimate the risk of extreme events. They are important for resilience planning and adaptation to climate change and are routinely produced using ensemble of forecasts from dynamical models.

We now have a first order understanding of the origin of the predictability on these long timescales and it often involves tropical mechanisms, where ocean atmosphere coupling is strong and long range predictability is high. In particular, tropical rainfall variations and associated atmospheric disturbances are highly predictable and can leak out into the extra-tropics, via planetary waves and other mechanisms to impart predictability to the extratropical flow (Scaife et al., 2017; 2019). However, the extratropical forecast signals contain large amounts of internal, unpredictable variability, necessitating the generation of large, computationally expensive ensembles to extract forecast signals. Furthermore, dynamical models are not perfect models and contain errors e.g., in the representation of extratropical teleconnections.

This project will use machine learning to investigate the potential for skilful extratropical long-range forecasts using only observed or predicted tropical rainfall. It will assess the skill of ML generated forecasts and compare them to state of the art dynamical model forecasts, as a potentially cheap and skilful method of ensemble forecast generation.

The project will address this topic by investigating the following research questions:

What is the skill of different forecast suites (ML vs. DynFC) with respect to core metrics of tropical and extra-tropical weather and climate? Special focus will be laid on the predictability of relevant large-scale modes (i.e., ENSO, PNA, NAO, etc.) and the predictability of extreme events (i.e., extra-tropical cyclones and storms, heat waves, extreme precipitation).
Which factors (e.g., tropical rainfall variability, ENSO, West Pacific High) cause predictability on different time scales and how are the relevant mechanisms (e.g., wave propagation to the extra-tropics) differently predicted?

For which geographical regions are skilful prediction most likely and how would they differ between ML- and dynamical-based forecasts?
How is predictability explained in terms of physical processes? This underpinning understanding is essential to give comfort to decision makers, and it gives the basis for insights into another key consideration i.e., what is the uncertainty in the predictions?

Case funding

This project is suitable for CASE funding

Host

University of Birmingham

Theme

Climate and Environmental Sustainability

Supervisors

Project investigator

Prof GC Leckebusch, University of Birmingham, [email protected]

Co-investigators

Prof Adam Scaife, UK Met Office

Dr Doug Smith, UK Met Office

How to apply

Each host has a slightly different application process.
Find out how to apply for this studentship.
All applications must include the CENTA application form.
Choose your application route

Methodology

Based on state-of-the-art dynamical model seasonal to decadal predictions, respectively the UK Met Office’s GloSea6 and DePreSys (REF), the added value of machine learning generated long-range forecasts will be assessed. Skill will be quantified with classical measures like Kendall correlations or receiver operating characteristics (ROC). Extreme events will be diagnosed by established, index-based approaches or for more dynamical features like European windstorms via a dedicated identification tracking algorithm (cf. Leckebusch et al., 2008).

Training and skills

Students will be awarded CENTA2 Training Credits (CTCs) for participation in CENTA2-provided and ‘free choice’ external training. One CTC equates to 1⁄2 day session and students must accrue 100 CTCs across the three years of their PhD.

Specifically, for this project, the PhD student will gain state-of-the-art data science skills applicable to very large meteorological data (magnitude of Petabytes). Training will be provided in meteorological and climatological data analysis techniques, diagnostic approaches used in atmospheric science, and suitable statistical methods (e.g., in R, Python) to quantify forecast skill objectively. To gain in-situ
experience the student will have the opportunity to undertake a placement at the Met Office during the course of the project focussing on seasonal- to decadal prediction generation and related production of forecasts for decision makers in different sectors of society, incl. the energy- or finance sector.

Partners and collaboration

The Met Office is the national meteorological service for the UK. They provide critical weather services and world-leading climate science, helping to make better decisions to stay safe and thrive. As a world leader in providing weather and climate services, The Met Office employs more than 1,700 people at 60 locations throughout the world. Recognised as one of the world’s most accurate forecasters, the Met Office uses more than 10 million weather observations, an advanced atmospheric model and a high-performance supercomputer to create 3,000 tailored forecasts and briefings every day. These are delivered to a huge range of customers from the Government to businesses, the general public, armed forces, and other organisations. Thus, the Met Office plays a key role on the international stage by providing vital services, advancing global understanding through research and being an important participant in projects and organisations.

Further details

Further details on how to contact the supervisor for this project and how to apply for this project can be found here:

For any enquiries related to this project please contact Prof GC Leckebusch, [email protected].

To apply to this project:

You must include a CENTA studentship application form, downloadable from: CENTA Studentship Application Form 2024.
You must include a CV with the names of at least two referees (preferably three) who can comment on your academic abilities.
Please submit your application and complete the host institution application process via: https://sits.bham.ac.uk/lpages/LES068.htm. Please select the PhD Geography and Environmental Science (CENTA) 2024/25 Apply Now button. The CENTA application form 2024 and CV can be uploaded to the Application Information section of the online form. Please quote CENTA 2024-B30  when completing the application form.

Applications must be submitted by 23:59 GMT on Wednesday 10th January 2024.

Possible timeline

Year 1

The student will familiarise themselves with state-of-the-art seasonal to decadal forecasting, including e.g., UK Met Office GloSea6 and DePreSys forecast suites. They will identifying candidate metrics and physical processes with promise to be skilful predictors for further investigations. In parallel, the student will familiarise themselves with current ML-approaches e.g., used for NWP-timescales or up to long-range forecasting.

Year 2

Hypothesis driven investigations of observable quantities (e.g., storm counts, heatwave intensities), and the physical processes driving predictability. Application of AI tools in multi-member ensembles. Quantification and detailed analysis of skill of the different approaches.

Year 3

Critical Analysis of the usability of ML-based predictions and their physical reasoning. Identification of regions of enhanced skilful predictions. Deriving guidelines of best-practice of potentially suitable ML-based predictions in cooperation with dynamical ensemble-based forecasting.

2024-B30 ​​Machine Learning for Long Range Forecasts​