Machine Learning Workshop

Tuesday 29 March

Machine learning for emulation of model components

Chair: Matthew Chantry

Tom Beucler

University of Lausanne

Physics-guided and causally-informed machine learning for climate modelling

By incorporating thermodynamics into machine learning models of storms and clouds, we can help the models generalise across a wide range of climates. But what if we lack the relevant physical knowledge for generalisation?

By selecting the most informative predictors via causal discovery we can alleviate the generalisation and stability problems that limit the potential of machine learning for climate modelling.

Paula Harder

Fraunhofer Institute ITWM

Physics-informed learning of aerosol microphysics

Despite their importance for climate processes, aerosols are often not modelled in detail in global models, due to high computational costs.

To tackle that problem, we train a fast physics-constrained neural network to emulate an aerosol microphysics model, achieving sufficient accuracy and a speed-up of 64x.

Wednesday 30 March

Machine learning for forecasts from now-casting to seasonal

Chairs: Peter Dueben, Shraddha Gupta and Nikolaos Mastrantonas

Petrina Papazek

ZAMG

Solar power forecasts using deep learning

Fostering solar power as a sustainable energy source demands accurate, location-optimised, and highly resolved forecasts of power production.

Deep learning enables us to consider data from historical time series to model highly variable changes in expected power production and so optimise forecasts based on machine learning.

David Landry

Computer Research Institute of Montreal

Choosing the best seasonal forecasting model using neural networks

As part of an international prize challenge to apply artificial intelligence to weather data, we combine established and well-known methods with newer neural networks to improve seasonal forecasts.

These long-range forecasts have applications in critical areas such as renewable energy production, natural resource management, and climate resilience in general.

Dale Durran

University of Washington, Seattle

Deep learning weather prediction: new scientific horizons

Deep-learning-based weather prediction (DLWP) models can make data-driven predictions for the evolution of our atmosphere in ways that go beyond our current physics-based understanding. What new horizons can this, and other unique properties of DLWP, open for scientific exploration?

Thursday 31 March

Machine learning for feature detection and user applications

Chairs: Mihai Alexe and Jesper Dramsch

Arnaud Mounier

Météo-France

Using a neural network to explore rainfall patterns

Extracting and conveying relevant information from ensemble forecasts can be challenging, but is vital for making accurate predictions about weather.

We introduce an approach that uses a neural network to extract information on rainfall patterns and check for similarities with previous events and so define predictability of individual events.

Clara Betancourt

Jülich Supercomputing Centre

How ozone research can benefit from explainable machine learning

Machine learning approaches are widely applied to predict the spatial variability of tropospheric ozone.

We show how the reliability and generalisability of machine learning models for ozone prediction can be evaluated by analysing their latent spaces and feature spaces. This allows us to gain new insights across the global domain.

Friday 1 April

Machine learning tools and high-performance computing

Chairs: Peter Dueben & Florian Pinault

Peter Bauer

ECMWF

The European Union Destination Earth project and its potential for boosting machine learning

Digital twins will advance how we create and use information that helps us better adapt to climate change and manage the risks of environmental extremes.

Machine learning will be key to achieving the speed and versatility that is needed for the unique level of interactivity these enhanced models of Earth will provide.

Bianca Zadrozny

IBM Research

Generative machine learning for extreme climate scenarios

There is growing need for plausible ‘what-if’ climate scenarios to help users understand exposure to extreme events, and examine resilience and mitigation strategies.

Machine learning techniques offer a pathway for generating climate scenarios and can improve the characterisation of ‘grey swan’ events; those with a low likelihood of happening, but a potentially large impact.

All the sessions are available to watch again via the event webpage.

We have an extensive education programme to support Member States and Co-operating States in the training of scientists in numerical weather forecasting, and in making use of our forecast products.

Find out more at ecmwf.int/en/learning