Agenda

Wednesday 2nd October 2019

Hackathon day

09:30 - 10:00 : Welcome coffee

10:00 - 10:45 : Schedule of the day and introduction to the challenge

10:45 - 11:00 : Team finalization - technical question

11:00 - 12:30 : Teamwork 1

12:30 - 13:30 : Snack

13:30 - 15:30 : Teamwork 2

15:30 - 16:00 : Pooling of progress

16:00 - 16:30 : Coffee break

16:30 - 18:00 : Teamwork 3

18:00 - 18:30 : Conclusions - next steps

Thursday 3rd October 2019

Workshop Day 1

08:30 - 09:00 : Registration and welcome coffee

09:00 - 09:15 : Welcome and intro

09:15 - 10:15 : Invited talk - Pascale Braconnot

Investigating long term climate variability and changes with Earth System models

10:15 - 11:15 : Invited talk - Michael Ghil

Low-Frequency Climate Variability: Markov Chains and Nonlinear Oscillations

11:15 - 13:15 : Lunch break

13:15 - 13:30 : Sponsor talk - Jennifer Marsman

Microsoft AI for Earth talk

13:30 - 14:30 : Invited talk - Marlene Kretschmer

Assessing teleconnection pathways with causal inference techniques

14:30 - 15:30 : Poster session #1 - Coffee served (see here the repartition of the posters)

15:30 - 16:30 : Invited talk - Pierre Gentine

Hybrid modeling: best of both worlds?

16:30 - 18:30 : Free time + optional labeling session for the ClimateNet project

18:30 - 21:00 : Reception

Friday 4th October 2019

Workshop Day 2

08:30 - 08:45 : Registration

08:45 - 09:45 : Highlight talks

08:45 - 09:05 : Eniko Szekely - A direct approach to detection and attribution of climate change

09:05 - 09:25 : Ibrahim Ayed - Learning the hidden dynamics of ocean temperature with Neural Networks

09:25 - 09:45 : Jussi Leinonen - Generative Adversarial Network for Climate Data Field Generation

09:45 - 10:15 : Coffee break

10:15 - 11:15 : Invited talk - Sebastian Engelke

Graphical models and causality for extreme events

11:15 - 13:15 : Lunch break

13:15 - 14:15 : Invited talk - Yoshua Bengio

AI and the Climate Crisis

14:15 - 15:15 : Poster session #2 - coffee served (see here the repartition of the posters)

15:15 - 16:15 : Highlight talks

15:15 - 15:35 : Christian Reimers - Using causal inference to globally understand black box predictors beyond saliency maps

15:35 - 15:55 : Pierrick Bruneau - Computing flood probabilities using Twitter: application to the Houston urban area during Harvey

15:55 - 16:15 : Jorge Baño-Medina - The importance of inductive bias in convolutional models for statistical downscaling

16:15 - 16:35 : Hackathon feedbacks

16:35 - 16:40 : Conclusions and final announcements, group photo

18:00 - 18:30 : Meeting point for the boat tour

18:30 - 19:45 : Boat tour

Invited talks

Pascale Braconnot

LSCE-IPSL, unite mixte CEA-CNRS-UVSQ, Université Paris-Saclay, France

Investigating long term climate variability and changes with Earth System models

Climate models, called Earth System Models, coupling ocean, atmosphere, land surface and sea-ice components though the energy, the water and biogeochemical cycles, have become key resources to understand how the climate works and evolved in response to natural or anthropogenic forcing. The sciences questions to be addressed now are not limited to the mean climate changes. They require being able to explore climate trends, variability or extremes at the global or regional scales, as well as the interactions between climate and the environment. In this presentation I will first provide an overview of some of the questions and the need for simulations with increased model complexity, improved resolution, longer integration and with several members. Then, using Holocene snap shot or transient (the last 6000 years) climate simulations I will illustrate some of the challenges behind long term simulations designed to understand climate feedbacks, the linkages between changes in the climate mean state and variability, or the ability to represent climate variations outside the modern range. This will also include some thoughts on model development and tuning, and the use of new methodologies, based on entropy and graph theory to guide the analyses.

Yoshua Bengio

Mila, Department of Computer Science and Operations Research, Université de Montréal

AI and the Climate Crisis

AI is moving out of universities and into society, which gives a new social responsibility to AI researchers. Climate change is the biggest crisis that humanity is currently facing. Machine learning is not guaranteed to help tackle this crisis – but it can help. From optimizing energy forecasting to synthesizing new molecules for batteries and improving crisis response, there is a plethora of ways in which ML can be used for both mitigation of and adaptation to the climate crisis. The presentation will discuss in more detail climate-related projects at Mila and elsewhere, including work on synthesizing new materials and on visualizing the effects of climate change.

Pierre Gentine

Department of Earth & Environmental Engineering, Earth Institute, Data Science Institute, Columbia University

Hybrid modeling: best of both worlds?

In recent years, we have witnessed an explosion in the applications of machine learning, especially for environmental problems. Yet for broader utilization, those algorithms may need to respect exactly some physical constraints such as the conservation of mass and energy. In addition, environmental applications (e.g. drought impact) are typically focusing on extremes and basically on out-of-sample generalization. This can be a problem for typical algorithms, which typically interpolate very well. I will here show how a hybridization of machine learning algorithms, imposing physical constraints within them, can help tackle those different issues and offer a promising avenue for environmental applications and process understanding.

Marlene Kretschmer

Potsdam Institute for Climate Impact Research, University of Reading

Assessing teleconnection pathways with causal inference techniques

Teleconnections refer to recurrent large-scale pressure patterns with low-frequency variability, connecting far-away geographical regions. They reflect modifications of atmospheric circulation affecting e.g. the position of the jet stream, storm track intensity or Monsoon strength and thus have a strong impact on our weather. However, extracting the physically relevant teleconnection pathways from observation or model data remains challenging. One major issue is to separate the signal from the noise given large internal atmospheric variability. This is compounded by varying dimensions in space and time and competing effects of different processes. Here, we discuss how novel data-driven causal methods beyond the commonly adopted correlation techniques can help to overcome some of these current limitations. We give an overview of causal inference frameworks and identify promising application cases common in climate science.

Michael Ghil

Ecole Normale Supérieure, Paris, and University of California, Los Angeles

Low-Frequency Climate Variability: Markov Chains and Nonlinear Oscillations

Two complementary ways of describing, understanding and predicting intraseasonal atmospheric variability have been proposed, episodic and oscillatory (Ghil & Robertson, PNAS, 2002). Recent progress in the methodology and results of these two approaches will be presented for subseasonal-to-seasonal (S2S) variability (Ghil et al., in Robertson & Vitart, Eds., Elsevier, 2018).

Sebastian Engelke

Research Center for Statistics, University of Geneva

Graphical models and causality for extreme events

Climate extremes such as heat waves, heavy rainfall or flooding attract an increasing attention by researchers and the public. The accurate statistical assessment of the small occurrence probabilities of such rare scenarios is based on extreme value theory. We will first give a short introduction to this theory and the well-established tools used for univariate data. Many practical questions however concern many variables at the same time. Climate scientists observe an increasing risk of compound events due to a combination of different variables, such as wildfires caused by low precipitation and extreme heat. Similarly, the flood risk of a river catchment depends highly upon the network structure and whether floods at different locations occur simultaneously or independently of each other. Recent advances in extreme value theory therefore concentrate on the dependence between rare events in complex multivariate or spatial systems.

Graphical models have recently been seen to be powerful tools for the analysis of such complex extreme events. We will present several methods to estimate underlying graph structures in a data driven way. This provides sparse and interpretable statistical models even in higher dimensions which can be easily communicated to practitioners. Directed graphical models are also the basis for causal inference. Causality for extremes is a hot topic at the moment with many applications, including the detection of causes for climate extremes. In the framework of linear structural equation models with heavy-tailed noise variables, we will present a computationally efficient algorithm to discover causal mechanisms that manifest in the extreme values of the data.

Jennifer Marsman

Microsoft, Microsoft AI for Earth

Microsoft AI for Earth

In this brief session, you will learn about Microsoft’s $50 million USD investment in AI for Earth grant funding, as well as an example of how a grant recipient is using this program to combat climate change.