This research proposal builds on the study by Koldunov, Nikolay, et al. (2024). and proposes to further use AI/ML techniques to effectively downscale climate model output using AI-based numerical weather prediction (AI-NWP) systems, such as the Pangu-Weather model or NVIDIA’s FourCastNet. These two models were trained on 40 years (1979-2018) of data from the ERA5 reanalysis, which combines numerous observations into a best estimate of the state of the atmosphere at a resolution of around 25 km. Such an AI-NWP model, when initialized with low-resolution initial conditions, has been shown to produce short-term forecasts with a higher level of detail corresponding to their training data.

For the current work, the initial phase involves downscaling from approximately 200 km to 25 km using AI for Numerical Weather Prediction (AI-NWP) techniques. We plan to use AI-NWP models to downscale low-resolution climate models without additional training, ensuring minimal computational costs and rapid progress. The objective is to evaluate the model’s capacity to produce physically plausible output fields with finer detail than the input data. If necessary, transfer learning will be employed to fine-tune the AI-NWP model using ERA5 and other relevant datasets to improve accuracy. We will also explore applying the AI-NWP model to low-resolution datasets such as CMIP6 (and CMIP7 when available), Geoengineering Model Intercomparison Project (GeoMIP), and Assessing Impacts and Responses of Climate Intervention on the Earth Systems (ARISE). Downscaled data from GeoMIP and ARISE will specifically be used to assess the impacts of Solar Radiation Management (SRM) scenarios. If the model’s performance does not meet the required accuracy, further refinement will be done using bias correction and/or transfer learning and relevant GeoMIP, ARISE, and ERA5 data.

Tasks

• Use AI/ML models, such as Pangu-Weather or NVIDIA’s FourCastNet, to improve the resolution of climate models.
• Work with large datasets, including ERA5, and Copernicus regional reanalysis for Europe (CERRA), for model training, downscaling, and fine-tuning.
• Apply advanced AI techniques, such as transfer learning, to adapt models for different climate scenarios (e.g., mid-21st century climates, SRM-cooled climates).
• Correct biases and enhance the accuracy of downscaled climate projections.
• Collaborate with global climate scientists, especially from the Global South, to ensure models meet diverse regional needs.
• Help build a user-friendly platform that enables users to input CMIP6,  GeoMIP, and ARISE data to obtain high-resolution downscaled results with visualization and analysis tools.

More

Duration
1+1 years for the first phase, with the possibility of extension up to 5 years

Salary
Net monthly salary around 2800€

Contact
Kazem Ardaneh •

• Advanced degree in computer science, engineering, or related fields.
• Expertise in Deep Learning techniques, preferably for numerical weather prediction or climate modeling.
• Experience with Python, TensorFlow and/or PyTorch, and large-scale data processing.
• Strong problem-solving skills and the ability to work collaboratively with interdisciplinary teams.

Desired

• Experience in bias correction and generative modeling to improve spatial resolution.
• Excellent communication skills to work in a collaborative research environment.
• Familiarity with climate models such as CMIP6, GeoMIP, and ARISE, with experience handling large climate datasets.