Machine Learning seminar series

Seminar series | Live-streamed

Machine Learning seminar image

Upcoming talks

Building trustworthy AI for environmental science | 8 September | 16:00 BST

Host

Baudouin Raoult (ECMWF)

Speaker

Amy McGovern is a Lloyd G. and Joyce Austin Presidential Professor in the School of Computer Science and an Adjunct Professor in the School of Meteorology at the University of Oklahoma. She has been leading the development of AI/ML for weather applications for 15 years.  

 

Abstract

As climate change affects weather patterns and sea levels rise, the world’s need for accurate, usable predictions of weather and ocean and their impacts has never been greater. At the same time, the quantity and quality of Earth observation and modeling systems are increasing dramatically, offering a deluge of data so rich that only automated intelligent systems can fully exploit it.  In this talk, I will discuss our approach to developing trustworthy AI methods for environmental science.  I will also present preliminary results for high-impact weather prediction and understanding.

From research to applications – Examples of operational ensemble post-processing using machine learning |1 October | 11:30 BST

Host

Zied Ben Bouallegue (ECMWF)

Speaker

Maxime TaillardatMaxime Taillardat received a M.Sc. Degree in computer science from the National Institute of Electrical engineering, Electronics, Computer science, Fluid mechanics & Telecommunications and Networks, Toulouse, France, a M.Sc. Degree in meteorology, statistics & machine learning from the National   Meteorology   School,   Toulouse,   France,   and   the   Ph.D.   degree   in   meteorology, oceanography and environmental science from Université Paris-Saclay, Versailles, France. He works in the Statistical Forecasting and Verification team in Météo-France, the French Weather Service. He is also affiliated to the National Centre for Meteorological Research – UMR 3589, Toulouse, France. His research interests include among others the use of machine learning algorithms in weather forecasting, especially for the post-processing of numerical weather prediction models, decision sciences, and the verification of ensemble forecasts for extreme events. Since 2018, he is one of the conveners of the session ''Advances in statistical post-processing for deterministic and ensemble forecasts'' at the European Geosciences Union General Assembly.

Abstract

Statistical post-processing of ensemble forecasts, from simple linear regressions to more sophisticated techniques, is now a well-known procedure in order to correct biased and poorly dispersed ensemble weather predictions. However, practical applications in National Weather Services is still in its infancy compared to deterministic post-processing. We present two different applications of ensemble post-processing using machine learning at an industrial scale. The first is a station-based post-processing of surface temperature in a medium resolution ensemble system. The second is a gridded post-processing of hourly rainfall amounts in a high-resolution ensemble prediction system. The techniques used rely on quantile regression forests (QRF) and ensemble copula coupling (ECC), chosen for their robustness and simplicity of training whatever the variable subject to calibration.

Moreover, some variants of classical techniques used such as QRF or ECC have been developed in order to adjust to operational constraints. A forecast anomaly-based QRF is used for temperature for a better prediction of cold and heat waves. A variant of ECC for hourly rainfall is built, accounting for more realistic longer rainfall accumulations. It is shown that forecast quality as well as forecast value is improved compared to the raw ensemble, which is critical in the context of forecast automation. At last, comments about model size and computation time will be done.

Causal Networks as a framework for climate science to improve process understanding | 27 October | 16:00 BST

Host

Inna Polichtchouk (ECMWF)

Speaker

Marlene KretschmerMarlene Kretschmer is a post doctorate researcher at the University of Reading. Before that she worked at the Potsdam Institute for Climate Impact Research in Germany where she received her PhD in climate physics. Her research focuses on the dynamical Stratosphere-Troposphere coupling and its impacts for winter circulation and especially for extreme weather events. To address these issues, she is particularly interested in applying novel statistical approaches from machine learning such as causal discovery algorithms.  Moreover, she is keen on applying these new techniques to evaluate teleconnection processes in climate models and to improve sub-seasonal to seasonal (S2S) forecasts.

Abstract

In the light of ongoing anthropogenic climate change and associated risks, supporting regional decision making should be a guiding principle of climate research. However, seasonal forecast models only have low skill and climate models often give inconclusive results about regional aspects of climate change. One major source of uncertainty are dynamical drivers in the climate system, such as storm tracks or blocking, which are not well understood theoretically and where models show diverse responses.

The recent hype of machine learning promises data-driven solutions to these issues. While data-centric methods such as deep learning have and certainly will make notable contributions to the earth sciences, their power lies in their ability to efficiently describe complex relationships present in the data. There is reason to doubt whether these methods can, on their own, deal with the sort of epistemic uncertainty described above. Moreover, machine learners and climate scientists often lack a common language, making successful collaboration still difficult. In particular, climate scientists are trained to think in terms of causal relationships, whereas machine learning is mostly descriptive (i.e. correlational) and does not explicitly incorporate domain knowledge.

Here we call for the use of causal networks in climate science as a framework to overcome some of these challenges. We argue that causal networks are a simple yet powerful tool to translate qualitative expert knowledge about physical processes into mathematical objects, to gain quantitative information about the role of these processes through applying the rules of causal inference.

Recordings and slides of past talks

Probabilistic downscaling to detect regional present and future climate hazards | 28 April

 

Speaker: Sherman Lo (University of Warwick)

Seminar recording

PDF icon Presentation slides

Exploring Machine Learning for Data Assimilation | 7 May

Seminar recording

Presentation slides

Speaker: Alban Farchi (ECMWF)

MetNet: A Neural Weather Model for Precipitation Forecasting | 12 May

Speaker: Nal Kalchbrenner

Seminar recording

Presentation slides

AI, a change in science/technology ... or culture? | 14 July

Speaker: Alberto Arribas

Seminar recording

Presentation slides

Spatiotemporal complexity and time-dependent networks in mid- to late Holocene simulations | 28 July

Speaker: Dr Annalisa Bracco
 

Seminar recording

Presentation slides

Other Machine Learning Talks at ECMWF

 

18 June
at 10:30 BST

VIRTUAL

Link to join the seminar

Title: "ecPoint” - a Post-processing Tool that improves Forecasts and highlights Systematic Model Errors

Speaker: Timothy Hewson

Abstract

In April 2019 ECMWF introduced a new, experimental, “point rainfall” forecast product onto its ecCharts web display platform, based on the post-processing package “ecPoint”, to give site-specific forecasts for everywhere in the world up to day 10. Prior to this development forecasters had access to just raw ensemble output in ecCharts, which provides gridbox average totals. ecPoint aims to incorporate probabilistically the expected sub-grid variability, and simultaneously apply gridscale bias corrections. Both these adjustments depend critically on “gridbox-weather-type”.

This presentation will describe the meteorology-based calibration rationale that underpins ecPoint, how this is different to pre-existing post-processing methods, and how it can also be applied to other surface variables such as 2m temperature. Numerous benefits will be highlighted.

The conditional verification concepts underpinning the calibration allow one to identify weather-situation-dependant gridscale biases. Examples will illustrate the diagnostic power of this approach, showing where and when rainfall is typically under- and over-forecast, providing pointers for future model improvements. And using an open source GUI one can apply the calibration code to data from other models, and thereby intercompare performance in different weather situations.

The forecast improvements that then arise will be discussed, using both long term global verification up to day 10, and illustrative case studies, with a focus on how extreme localised rainfall, that might lead to flash floods, is better handled. It will be shown how the post-processing can usefully shift the emphasis for warning issue from one region to another, when one compares with raw ensemble output.

There will be brief reference, from collaborative work, to how ecPoint output seems to compare favourably with the post-processed output of convection-resolving limited area ensembles.

The talk will conclude by discussing, in the context of ongoing and potential projects, numerous future applications of ecPoint, such as bias-corrected inputs to hydrological models, point rainfall re-analyses and tests of theories such as city impact on rainfall. Avenues for improving the methodology will also be highlighted.

Attendance

The links to join the seminars will be published on this page. If you would like to receive the joining instructions for each seminar by email, please complete the registration form.

Registration

Contact


#AIseminars2020