Detailed programme of the MOOC
Tier 2 - Concepts of Machine Learning
This tier focuses on the key concepts of Machine Learning. You will learn how to apply Machine Learning techniques relevant to weather and climate applications. You will also gain hands-on experience in selecting and designing Machine Learning workflows and applying them to simplified real-world problems.

Introduction to data handling
Launch date: Monday, 20 February 2023
Duration: 3h e-learning and 1 webinar
Programme:
The vast majority of time spent by data scientists is on the preparation of data before it can be fed into a Machine Learning workflow. This preparation may include handling missing or incomplete data, structuring data into a consistent format, or feature engineering, which is the creation of additional variables (features) to your dataset to improve machine learning model performance and accuracy.
This module will cover the main steps involved in this process for typical weather and climate datasets. It will provide guidance in best practices and include exercises to read data into Python objects and manipulate them with common Python libraries.
We will discuss data splitting and the possible pitfalls when learning on spatial-temporal data.
Expert: F. Pinault (ECMWF)

Regression and decision trees
Launch date: Monday, 27 February 2023
Duration: 2h15 e-learning and 1 webinar
Programme:
Regression and decision tree techniques strike a balance between being powerful learning tools whilst being interpretable. For example, decision tree methods often outperform deep learning techniques on tabular data.
In this module we will unpack how each of these methods work, including how to empower regression techniques with nonlinear features. We will introduce the toolboxes that make using these techniques easy.
Expert: M. Chantry (ECMWF)

Deep learning architectures
Launch date: Monday, 6 March 2023
Duration: 2h15 e-learning and 1 webinar
Programme:
The modern-day machine learning boom was caused by deep learning. Advances in many fields especially those covered in the media, use deep neural networks. The main advantage of deep learning is its versatility and composability, where we develop purpose-built architectures from nowcasting to long-term climate prediction.
These neural architectures are composed of building blocks in machine learning. We will cover architectures such as convolutional and recurrent neural networks, from long short-term memory networks to modern state-of-the-art transformers and discuss their strengths and limitations in analysing complex weather and climate data. We will also delve into the small details that make neural networks work in the real world of weather and climate prediction. By the end of this lesson, learners will have a good overview of the different deep learning architectures and their potential for improving weather and climate prediction.
Expert: J. Dramsch (ECMWF)

Uncertainty & generative modelling
Launch date: Monday, 13 March 2023
Duration: 1h45 e-learning and 1 webinar
Programme:
Many standard machine learning methods are criticised by users because they lack the ability to express uncertainty. However, knowing the uncertainty of a prediction is crucial in the field of weather forecasting. At most weather centres, this uncertainty is currently quantified by running an ensemble i.e., running a model multiple times with perturbed initial conditions.
This module will explore techniques to quantify uncertainty in standard machine learning methods and look at how novel machine learning methods can make probabilistic predictions. It will also cover generative models and show how they can be used to generate an ensemble of realistic weather predictions. In particular, this module will include Bayesian Neural Networks and Generative Adversarial Networks (GANs).
Expert: M. Clare (ECMWF)

Physics-Guided Machine Learning
Launch date: Monday, 20 March 2023
Duration: 2h45 e-learning and 1 webinar
Programme:
Physics-Guided Machine Learning restricts the output of machine learning algorithms to physically-plausible solutions using physical information.
In this module, you will learn about common advantages, limitations, and state-of-the-art applications of physics-guided machine learning for weather and climate predictions. We will guide you through the key steps and challenges of formulating a physics-guided machine learning framework via two hands-on exercises: Physically-constrained neural networks to postprocess surface weather, and physically-informed parameterization of subgrid-scale tendencies for climate modeling.
Expert: T. Beucler (University of Lausanne)