Weather forecasts beyond the intrinsic weather timescale of 10 days or so, but for timescales shorter than seasonal means, are an exciting new area of research. Such ‘sub-seasonal to seasonal (S2S)’ forecasts, especially of heat waves or of increased likelihood of storms, are of great relevance for decision-making in several sectors such as renewable energy, health and agriculture. However, forecasts on these lead-times are usually strongly limited in skill and interpretability, and therefore in their useability. This PhD project aims to improve S2S predictions of the likelihood of extreme weather events through developing and applying hybrid statistical-dynamical methods.
S2S predictions are currently made with large ensembles of physics-based climate models. Although these dynamical models are capable of simulating the Earth system reasonably well, they often give inconclusive results about the regional occurrence of extreme events. One key issue is that atmospheric teleconnection pathways affecting regional weather patterns are not well represented in these models. For example, recent work showed that while dynamical models simulate the Madden-Julian Oscillation (MJO) – a key tropical driver of S2S weather anomalies world-wide – well, the extratropical response to active MJO phases is underestimated. In other words, there is untapped potential in dynamical climate models emerging from large-scale remote drivers such as the MJO.
To fill this gap, data-driven models offer promise. Data-driven prediction models are based on patterns found in past data. Methods from machine learning, and deep learning in particular, have been shown to be especially effective in detecting complex non-linear relationships in large data sets, and their application is becoming more and more important for weather and climate forecasting. However, while novel machine learning algorithms may lead to more accurate forecasts, the results are usually not physically interpretable and therefore generally not considered reliable and trustworthy within the weather and climate prediction community.
This PhD project will develop hybrid forecasts that combine predictions from dynamical models with data-driven, deep learning prediction models to optimally exploit the S2S sources of predictability. First, windows of opportunity for such a hybrid approach will be identified and tested. Next, deep learning methods will be applied to train and test prediction models during these windows of opportunity based on historical data. The resulting information will then be combined with the ensemble forecasts from the dynamical model using Bayesian methods to create a hybrid model or calibrated forecast. In other words, the predictions from the dynamical models will be boosted with background information using deep learning models.
Special focus will be placed on the physical interpretability and plausibility of the results. To this end, the hybrid forecasts will be embedded within a causal inference approach. This will ensure a physically consistent narrative of updated forecast information in the presence of different relevant and partly competing teleconnections. For instance, as well as the MJO, the stratospheric polar vortex is known to be a major driver of extreme European winter temperatures, and the impacts of the two drivers can either work in tandem or offset each other. The use of causal inference theory will ensure that the different contributions are disentangled and quantified accordingly, controlling for possible confounders. Moreover, methods from explainable AI will be applied to reveal what the deep learning models have learned from the data. This is important to overcome the black-box nature of these models and to build trust in their predictions.
The project will not only support better decision-making in weather-affected sectors but will also make a major contribution to the evolving field of physics-based machine learning.
Training opportunities:
- Visiting placements at ECMWF, to interact with experts on S2S prediction and machine learning. This will provide the opportunity for immersion into the environment of a leading operational centre, where the student can be infused with knowledge on predictability, diagnostics and databases.
- Regular visits to the Machine Learning Department at TU Berlin, where Marlene Kretschmer is collaborating with the group working on understandable ML.
- A 10-week placement with the Red Cross Red Crescent Climate Centre, to extend the methods developed in this project into the decision-making context.
For further details please contact [Email Address Removed].
Eligibility requirements: Applicants should hold or expect to gain a minimum of a 2:1 Bachelor Degree , Masters with Merit, or equivalent in meteorology, physics, mathematics or a closely related environmental or physical science. Strong computer and programming skills will be an advantage. Due to restrictions on the funding this studentship is open only to citizens of ECMWF member states. https://www.ecmwf.int/en/about/who-we-are/member-states
How to apply: please click the link here to apply. Create an account and during the application process please select the PhD in Atmosphere, Oceans and Climate. Please quote reference GS22_003 in “Scholarship applied for” box. When you are prompted by the online application system to upload a research proposal, please omit this step as the project is already defined.
Continue with Facebook
