Published: 22 April 2024
Lausanne
100%
Unlimited employment
Université de Lausanne
Job ads found online
UNIL is a leading international teaching and research institution, with over 5,000 employees and 17,000 students split between its Dorigny campus, CHUV and Epalinges. As an employer, UNIL encourages excellence, individual recognition and responsibility.
Recent progress in computer vision has accelerated the development of statistical downscaling, which uses statistical models to improve the spatiotemporal resolution of impactful climate variables, such as extreme temperatures, wind gusts, and precipitation. Machine learning (ML)-based super-resolution algorithms, which learn from data how to best generate high-resolution images from their low-resolution version, are gaining traction because of their improved accuracy and low omputational cost once trained. However, they are rarely designed to perform well on extremes, and their robustness is usually only tested in the present climate, where training data are available. These limitations prevent the widespread adoption of modern ML to better constrain uncertainties in the forecasting of local extremes and in the high-resolution projections of climate change.
To address these limitations, our project leverages recent developments in deep learning and geostatistics1 to design hybrid statistical-physical methods helping ML frameworks generalize to climate change and extreme events. We will test these methods for two applications that combine downscaling with prediction over Switzerland: future climate projections and medium-range forecasting.
For that purpose, we propose two synergistic PhD projects on Machine Learning for Weather/Climate Super-Resolution :
The first project generates spatially-resolved Swiss climate change scenarios that ensure physical consistency and preserve long-term trend.
The second project downscales medium-range forecasts over Switzerland, including small-scale features ignored by the original forecast.
In both projects, we will explore the added value of state-of-the-art ML for atmospheric science, which remains challenging to understand.
Both PhD positions will be hosted within UNIL’s recently established Expertise Center for Climate Extremes, and will include collaboration opportunities with research groups from the institutions listed below:
● MeteoSwiss
● LSCE-IPSL
● University of California, Irvine
● NVIDIA (Climate Simulation Team)
Expected start date in position : 01.10.2024 / to be agreed
Contract length : 1 year, renewable to a maximum of 4 years.
Activity rate : 100%
Workplace : Lausanne Mouline (Géopolis)
Technical Responsibilities:
Specific to the climate projection project:
● Develop and apply multivariate bias-correction methods suitable for the Swiss climate scenarios, focusing on enforcing physical consistency, maintaining inter-variable consistency, and evaluating spatial properties across downscaled outputs.
● Integrate and test new algorithms for trend preservation in the downscaled variables, ensuring that long-term climatic changes are accurately reflected in the projections.
● Collaborate with MeteoSwiss to ensure that the downscaled models align with the ongoing updates to Swiss climate change scenarios and are informed by the latest observational data.
Specific to the weather forecasting project:
● Design and implement machine learning models for medium-range weather forecasting that can handle initial conditions significantly different from those seen in the training data, especially under extreme weather conditions.
● Ensure the calibration of downscaled outputs for longer lead times as predicted distributions shift towards climatology.
● Develop and test feature transformation techniques that enhance the generalizability and accuracy of forecasts across complex topography, such as the Alpine region.
● Collaborate with MeteoSwiss to ensure that the downscaled forecasts align with ongoing forecasting and post-processing needs.
For both projects:
● Establish the broad applicability of new machine learning frameworks (new architectures, techniques to improve extrapolation/explainability, improved conditioning of the predicted distributions, etc.) for climate applications. You will receive support from the project’s team and have the opportunity to co-supervise Master's interns.
● Over the course of 4 years, complete 3-4 semesters of teaching assistantship in geoinformatics, scientific programming, or machine learning for Earth and environmental science courses (no more than approximately 5 hours/week).
● Given that the collaboration with MeteoSwiss offers the opportunity to root your research in practice, we encourage flexibility between Lausanne and the MeteoSwiss offices (Zürich, Geneva, Locarno).
Prof. Tom Beucler
tom.beucler@unil.ch
Deadline : 31.05.2024
Please, send your full application with all the following in PDF.
Only applications through this website will be taken into account.
We thank you for your understanding.
We are dedicated to fostering a diverse, equitable, and inclusive environment where all individuals are encouraged to apply, regardless of their background. UNIL is committed to equal opportunities and diversity.
www.unil.ch/egalite
UNIL supports early career researchers.
www.unil.ch/graduatecampus
The Faculty of Geosciences and Environment of the University of Lausanne adheres to the DORA agreement and follows its guidelines in the evaluation of applications (in short, quality over quantity)