Detailed programme of the MOOC

Tier 3 - Practical ML applications in Weather & Climate

Here you will explore the state of the art Machine Learning research throughout the key applications of numerical weather and climate predictions. You will gain hands-on experience with the use of code and data in real research projects. Finally, you will develop a case study in at least one of the main topics of numerical weather and climate prediction, by applying the practical application of Machine Learning workflows.


Launch date: Monday, 27 March 2023
Duration: 3h30 e-learning and 2 "Meet the Expert" Webinars
1. ML-based retrieval techniques for current and future satellite missions for weather and climate
This unit illustrates applications of machine learning techniques to support the development of next satellite missions and to retrieve weather and climate variables exploiting large observational datasets. We will focus on deep-learning applications for end-to-end simulations of missions measuring Thermal Infra-Red (TIR) spectra for land surface temperature retrievals developed in an ESA project, as well as machine learning techniques for precipitation retrieval from passive microwave and VIS/IR radiometer measurements developed within the EUMETSAT Satellite Application Facility for Hydrology (H SAF) in preparation for the future Meteorological Operational-Second Generation (MetOp-SG) satellite mission and the upcoming Meteosat Third Generation (MTG) mission.
Expert: G. Panegrossi (ISAC-CNR)

2. Machine learning for environmental modelling
This unit will survey the applications of machine learning methods in different areas of environmental science and evaluate how machine learning algorithms link remote sensing information to relevant environmental variables, taking the complexity and non-linearity of nature into account.
Expert: S. El Garroussi (ECMWF)

Forecast Model

Launch date: Monday, 27 March 2023
Duration: 3h e-learning and 1 "Meet the Expert" Webinar
1. AI for Nowcasting
The nowcasting forecast covers a time horizon of a few hours from the last acquired observations. The algorithms developed for this type of forecast take advantage of all available high time-frequency observations. The topics covered in this unit will be:

  • Use of observations in nowcasting
  • Definition of nowcasting algorithms
  • Limits and perspectives in the use of deep learning nowcasting techniques
  • Operational application and comparison with nowcasting techniques based on extrapolation
Expert: V. Poli (ARPAE)

2. Parametrisation emulation
Processes that cannot be explicitly simulated in Earth system models must be parameterized, which is difficult and leads to large uncertainties in predictions. In this unit, you will learn how machine learning can accelerate models by emulating existing parameterizations, or even improve models by directly learning parameterizations from high-fidelity data. We will discuss state-of-the-art approaches, their challenges, and how they may address shortcomings of traditional parameterizations.
Experts: M. Chantry (ECMWF) and T. Beucler (University of Lausanne)

Data Assimilation

Launch date: Monday, 3 April 2023
Duration: 2h e-learning and 1 "Meet the Expert" Webinar
Data assimilation, as described in Tier 1, is a way to combine the forecast of a numerical model with observations to improve the “state” of the Earth system. Combining machine learning and data assimilation makes it possible to go one step further and improve the forecast model itself. The output of the algorithm is a hybrid model that combines a physics-inspired part and a statistical part learned from observations.
What will be accomplished in this module:

  • Perform data assimilation on a toy model
  • Train a neural network to construct the data-driven part of the model
  • Demonstrate the improved forecast skill of the newly created hybrid model
Experts: J. Brajard (NERSC) and M. Bocquet (École des Ponts ParisTech)


Launch date: Monday, 3 April 2023
Duration: 1h e-learning and 1 "Meet the Expert" Webinar
We will apply Generative Adversarial Networks (GANs) to the practical problem of precipitation downscaling. We will start with low-resolution precipitation fields and use a GAN to increase the resolution of the images while still producing realistic-looking precipitation. We will then implement metrics for evaluating the quality of the GAN outputs.
Expert: J. Leinonen (MeteoSwiss)

Ocean & Climate

Launch date: Tuesday, 11 April 2023
Duration: 3h e-learning and 2 "Meet the Expert" Webinars
1. Unsupervised learning to understand the ocean
The ocean is vast with emerging patterns and behaviours. Scientists seek to find patterns to guide exploration. Machine learning is becoming an important tool for data exploration. Unsupervised learning is a good choice as we are often interested in what structures reside in the data.

This part of the module covers different types of unsupervised machine learning methods to explore the oceans. It also discusses model selection and validation, the fundamentals of clustering for physics, in particular in dynamical regimes, and provides an example case study.
Expert: M. Sonnewald (Princeton University)

2. European Extreme Events Climate Index (E3CI)
The European Extreme Events Climate Index (E3CI) is an ensemble of indices that provide information about the areas affected by different types of weather-induced hazards and the severity of such events. E3CI includes five components returning information about main hazards: cold and heat stresses, droughts, extreme precipitations, extreme winds.
The module will cover:

  • An explorative data analysis of E3CI components
  • Identification of trends in climate anomalies for specific regions of Europe
  • Use of clustering techniques to identify similar regions from a climatic point of view
Experts: A. Tirri, F. Lo Conti (Leithà) and G. Rianna (CMCC)

Operational Meteorology

Launch date: Tuesday, 11 April 2023
Duration: 2h00 e-learning
Operational meteorologists have had limited opportunities to learn about machine learning (ML). While traditionally this has not been a problem, more and more ML forecasting tools are being created annually. This module summarizes two manuscripts and code that were created to introduce ML written for operational meteorologists.

The module also provides an example of how machine learning can be used to increase the efficiency of operational tasks, such as resource scheduling.
Experts: B. Raoult (ECMWF) and R. Chase (University of Oklahoma)

Download the programme

< See Tier 2