SUMMER SCHOOL: ARTIFICIAL INTELLIGENCE FOR DETECTION AND ATTRIBUTION OF CLIMATE EXTREMES

MINUTES OF THE LECTURES

DAVIDE FARANDA – “PHYSICS LEARNING FOR THE ATTRIBUTION OF ATMOSPHERIC EXTREME EVENTS”

The first day of the summer school I presented an analogues-based methodology to perform extreme events attribution to climate change: this methodology is largely based on the use of dynamical systems theory which prescribes a precise mathematical way to look at the recurrences of a state in phase space (the space of all the variables of the system). This formalism is particularly useful in climate sciences because it allows to go beyond the detection of events similar to the one to attribute to climate change based on thresholds exceedances, it allows to perform attribution conditioned to a circulation leading to a specific extreme event. During the presentation I’ve also presented an application to the attribution of the Medicane Apollo, which is one of the XAIDA case’s study. At the end of the presentation, there was time to interact with the students that were very active in asking questions and suggesting also new line of research.

MIGUEL-ÁNGEL FERNÁNDEZ-TORRES – “SPATIO-TEMPORAL DEEP NEURAL NETWORKS FOR EXTREME EVENT DETECTION”

A review of the different deep learning methodologies for spatial, temporal and spatio-temporal modelling was done during the presentation, including Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). These methodologies were introduced with the goal of extreme event detection, illustrated by following a case study about drought monitoring during the severe Russian heat wave in 2010. Moreover, the talk explained the different concepts applied during Hackathons 2 and 3 (e.g. convolutional and pooling layers, activation functions, optimization, backpropagation, etc.).

ERICH FISCHER – “QUANTIFYING AND UNDERSTANDING VERY RARE EXTREMES”

The keynote lecture on “Understanding and quantifying changes in very rare extremes” by Erich Fischer demonstrated that recent weather and climate extremes broke long-standing observed records by large margins. Many of these unprecedented extremes often led to very large socio-economic and ecological impacts due to a tendency to adapt at most to the highest intensities experienced during a lifetime or documented historically. This lecture discussed challenges and opportunities of understanding and quantifying record-breaking events and different methods to assess the potential for future events unseen in the observational record. The lecture reviewed different methodological approaches to produce physical storylines and different lines of evidence from model-based storylines and physical process understanding.

EMANUELE BEVACQUA – “COMPOUND WEATHER AND CLIMATE EXTREME EVENTS”

Compound weather and climate events are combinations of climate drivers and/or hazards that contribute to societal or environmental risk. Studying compound events often requires a multidisciplinary approach combining domain knowledge of the underlying processes with, for example, statistical methods and climate model outputs. Recently, to aid the development of research on compound events, four compound event types were introduced, namely (a) preconditioned, (b) multivariate, (c) temporally compounding, and (d) spatially compounding events. In this talk, I introduced compound events and guidelines on how to study different types of events. The talk was based on a paper where we consider four case studies, each associated with a specific event type and a research question, to illustrate how the key elements of compound events (e.g., analytical tools and relevant physical effects) can be identified. Furthermore, examples of studies on projections of other types of compound events were given. Overall, the insights of the presentation can serve as a basis for compound event analysis across disciplines and sectors.

MATHIEU VRAC – “STATISTICAL DOWNSCALING AND BIAS CORRECTION”

The keynote lecture on “Statistical downscaling and Bias correction of climate simulations” by Mathieu Vrac first presented the needs for post-processing climate simulations before investigating climate change impacts. This lecture reviewed the various categories of statistical methods for downscaling (SD) and/or bias correction (BC), including discussions about multivariate aspects and methods.
Challenges and opportunities for developments and applications regarding extreme events were also discussed. The question of the stochasticity in SD and BC results, as well as potential developments regarding “physically-based” SD or BC methods (e.g. via large-scale circulation inclusion and modelling) were also parts of the presentation and discussed with the audience.

SEBASTIAN SIPPEL – “DYNAMICAL ADJUSTMENT AND DISTRIBUTIONAL ROBUSTNESS FOR D&A”

Abstract: Internal climate variability fundamentally limits short- and medium-term climate predictability, and the separation of forced changes from internal variability is a key goal in climate change detection and attribution (D&A). In this talk, I discuss the suitability of incorporating statistical learning techniques to identify forced and internal climate signals from spatial patterns of climate variables.
First, I demonstrate that statistical learning techniques can be used to characterize and remove the influence of internal atmospheric variability on target variables such as temperature or precipitation (“dynamical adjustment”), which enables a better characterization of the forced thermodynamic response of climate change. It is shown that well-known climate features such as the late 1980’s abrupt climate change in Switzerland can be explained by such techniques as a combination of internal variability and externally driven changes.
Second, I introduce a detection approach using climate model simulations and a statistical learning algorithm to encapsulate the relationship between spatial patterns of daily temperature and humidity, and key climate change metrics such as annual global mean temperature. Observations are then projected onto this relationship to detect climatic changes, and it is shown that externally forced climate change can be assessed and detected in the observed global climate record at time steps such as months or days. I discuss how these approaches can be extended to address key remaining uncertainties related to the role of decadal-scale internal variability (DIV). DIV is difficult to quantify accurately from observations, and D&A requires that models simulate internal climate variability sufficiently accurately. I show that recently developed statistical learning techniques, inspired by causality, allow to identify the externally forced global temperature response, while increasing the robustness towards different representations of DIV (via an explicit robustness constraint). The fraction of warming due to external factors, based on these optimized patterns, is more robust across different climate models even if DIV would be larger than current best estimates.

JAKOB RUNGE – “CAUSAL INFERENCE FOR CLIMATE SCIENCE”

The presentation of WP4 at the Summer School, held by Jakob Runge, was a one-hour presentation about the general framework of causal inference with a focus on applications to climate science and especially extreme events. The causality framework can answer causal questions in a setting where randomized experiments are not possible. This is particularly the case in climate science, where most data are observational or from simulations. In the talk, Jakob presented why and how causal inference methods have the potential to deepen the understanding of extreme events. The causality toolbox, which was introduced in the talk, can help find drivers of extreme events by going beyond mere association. However, using such methods on climate data brings a set of challenges. Among these are the high dimensionality of the data, non-linear dependencies, or unobserved variables. These and more were also presented in the talk. A short discussion followed after the talk, where attendees posed questions about specific use-cases where causality could be used.

AGLAÉ JÉZÉQUEL – “EXTREME EVENT ATTRIBUTION”

This presentation was two-part. The first part explained the different techniques used to perform extreme event attribution. The class covered topics such as : how to define the event of interest, the concepts of counterfactual and factual worlds, storyline and risk-based approach, the attribution of impacts (beyond weather hazards), and compound events attribution. The second part focused on the uses and usefulness of extreme event attribution. It was the occasion to discuss with the students about the challenges of interacting with stakeholders, and of how hard it is to evaluate the added value of scientific information in decision making. The discussion with the room was lively.

GUSTAU CAMPS-VALLS – “PHYSICS-AWARE MACHINE LEARNING”

Most problems in Earth sciences aim to do inferences about the system, where accurate predictions are just a tiny part of the whole problem. Inferences mean understanding variables relations, deriving models that are physically plausible, that are simple parsimonious, and mathematically tractable.
Machine learning models alone are excellent approximators, but very often do not respect the most elementary laws of physics, like mass or energy conservation, so consistency and confidence are compromised. I will review the main challenges ahead in the field, and introduce several ways to live in the Physics and machine learning interplay. Physics-aware machine learning models are just a step towards understanding the data-generating process, for which causality promises great advances. I’ll review some recent methodologies to cope with it too. This is a collective long-term AI agenda towards developing and applying algorithms capable of discovering knowledge in the Earth system.
https://arxiv.org/pdf/2010.09031.pdf
https://arxiv.org/abs/2104.05107

DIM COUMOU – “SPATIALLY COMPOUNDING SUMMER EXTREMES”

On the morning of Friday 24th of June, Dim Coumou gave a 1-hour lecture on spatially compounding summer extremes, including simultaneous heat waves in different regions in the northern hemisphere mid-latitudes. Such simultaneous extremes, driven by large-scale waves in the jet, provide substantial risks for global food production as the affected regions are important breadbaskets (US mid-west, western Europe, Russia/Ukraine and east Asia). Coumou introduced the different hypothesized drivers of quasi-stationary waves in the summer jet. Next, he presented a series of studies using both climate models and data-driven methods (in particular causal discovery analyses) to quantify the relevant importance of these hypothesized drivers, and how anthropogenic global warming (AGW) can affect those. He showed that AGW to date has likely led to a weakening of the summer jet and storm tracks, something that is also well-captured by the latest generation of climate models (CMIP6). However, there is still substantial uncertainty into how AGW affects quasi-stationary waves. Coumou highlighted the usefulness, and limitations, of causal discovery methods in teleconnection and extreme weather research.

SIMON KLEIN – “CLIMATE CHANGE EDUCATION – CO-WORKING ON SCHOOL ACTIVITIES AND PROJECTS”

As required by the climatologist community and IPCC experts as well as stated in the Paris Agreements, education is a key element to tackle climate change and take collective climate action. The Office for Climate Education (OCE), a UNESCO category 2 center, affiliated with Sorbonne University and IPSL is a major actor in contributing to climate change education via producing high quality pedagogical resources, and facilitating teachers training around the world.
OCE is one of the XAIDA WP2 members, and along with the MET Office, is contributing to the dissemination of XAIDA findings and research by creating sets of lesson plans, short videos and teachers training sessions on extreme events, involving the XAIDA community. In Trieste, Simon Klein presented the engagements of the OCE contributions to the XAIDA project, and the importance of working with young scientists in design pedagogical tools, but also to acknowledge, in the process, the importance of emotions that such results can trigger to the learners, such as the eco-anxiety.

DAVIDE FARANDA – “CHALLENGES IN DETECTION & ATTRIBUTION OF TROPICAL AND MEDITERRANEAN CYCLONES”

On the first day of the second week I present the challenges in attribution of tropical and mediterranean cyclones to climate change: indeed in recent years, multiple groups started to pursue attribution studies for specific convective events mostly focusing on tropical cyclones or tropical cyclone seasons. The goal of attribution is to evaluate the role of climate change in modifying the intensity and frequency of specific extreme weather events. The students were particularly interested in the limitations previously in observing and simulating tropical cyclones which prevent from directly attributing a severe cyclone to climate change: this would ideally require two large sets of high-resolution cyclones simulations with and without climate change. We had several follow-up discussions on this topic.

DAVIDE FARANDA – OUTREACH GAME “CLIMARISQ”

On the first day of the second week we also played ClimarisQ. ClimarisQ is a smartphone/web game from a scientific mediation project that highlights the complexity of the climate system and the urgency of collective action to limit climate change. It is an app-game where players must make decisions to limit the frequency and impacts of extreme climate events and their impacts on human societies using real climate models. The students gladly play the game and made several observations to improve its dynamics: most of their comments are already implemented in a new version of the game that should be out at the beginning of next year.

TED SHEPHERD – “BRIDGING PHYSICAL HYPOTHESES AND THEIR STATISTICAL ANALYSIS THROUGH CAUSAL NETWORKS”

Ted Shepherd gave a lecture prepared by Marlene Kretschmer entitled ‘Bridging physical hypotheses and their statistical analysis through causal networks’. The lecture introduced basic aspects of causal inference, meaning the quantification of known causal relationships, illustrated through application to atmospheric teleconnections. Causal inference is to be distinguished from learning causal relationships from data, known as causal discovery. Causal inference can be applied where there is already a large amount of expert knowledge available concerning physical hypotheses. This is the case with atmospheric teleconnections, which have a strong influence on the statistics of extreme weather and climate events. The scientific challenge is that of quantifying the causal linkages in the presence of multiple interacting factors, non-linearity, and non-stationarity. This approach also allows for a meaningful common language between dynamicists and statisticians, which is important for combining different lines of evidence. The lectures illustrated basic concepts such as blocked and open pathways of influence, common drivers, mediators, and common effects, and how these determine how to control for confounding influences in order to properly quantity causal effects. The students found the lectures very clear and pedagogical.

ROBERT VAUTARD – “STATISTICAL PRACTICE IN EXTREME EVENT ATTRIBUTION”

Robert Vautard gave a lecture explaining the principles of probabilistic event attribution, which rely on defining a class of events, such as the occurrence of a threshold exceedance, and analyzing whether such an event would have more or less probability to occur in a counterfactual climate altered by human activities. He explained also how such probabilities can be calculated from statistical modeling, and then followed by describing how such an approach is handled in operational attribution studies such as made in the “world weather attribution” network. He went on describing the protocol underpinning these activities, which allows a quasi-operational approach. He then provided several examples of event attribution analyses and their interpretations: extreme rainfall over the Seine river basin, drought in Madagascar, wind storm in Western Europe, growing-period frosts in France, Pacific North American heat wave. He then developed one of the main pending issues: why do models appear to underestimate heat waves in Western Europe? The question of model evaluation is absolutely crucial for attribution.
He concluded by describing main issues and bottlenecks when carrying out rapid attributions.

TED SHEPHERD – “STORYLINES”

Ted Shepherd gave a lecture entitled ‘The storyline approach to the construction of useable climate information at the local scale’. Storylines are self-consistent articulations of past or plausible future events or pathways, based on causal reasoning. They can be regarded as instantiations of the range of possibilities represented in causal networks, and thus are closely related to causal inference.
Storylines are particularly useful for representing combinations of epistemic (or systematic) and aleatoric (or probabilistic) aspects of uncertainty, which are not describable using standard (i.e. frequentist-based) statistical methods still widely used in climate science, such as null hypothesis significance testing. The lecture covered a range of aspects including the motivation for storyline approaches to regional climate risk, different kinds of storylines, the links to causality, the pitfalls of aggregation, the benefits from conditionality, and the interpretation of storylines within Bayesian statistics. In the days following the lecture, Ted had one-on-one discussions with many of the summer school participants. The connection with probability was perhaps the most common topic of discussion. Probabilities are appealing, but coming to terms with what are fundamentally subjective probabilities – as epistemic uncertainties necessarily are – is difficult for climate scientists.

ERIKA COPPOLA – “HEAVY PRECIPITATION EVENTS AND CONVECTION PERMITTING MODELLING”

The definition and concept of extreme precipitation has been introduced and several example presented to illustrate how the definition of extreme precipitation may be different in space and time and also how different can be the consequences in term of impacts. A brief recap of what is convection has been presented and which are the main limitations in modelling convection in the state of the art climate models where this process is parametrised and how this is relevant for the uncertainty of climate model projection. Therefore why the increase in resolution is needed to be able to explicitly model the convection and the state of the art of convection permitting regional climate modelling technique has been presented. Which are the tools and methodologies for this kind of dynamical downscaling, like for example which equations are needed for the model core since the hydrostatic approximation cannot be used anymore, which are the microphysical scheme that can be used for an explicit representation of convection (5-class microphysics for example), which are the possible nesting techniques to be used.
The need of very high computational resources and storage space has been explained for such kind of simulations and the discussion on the limitation of the model physics introduced. A summary of the main finding was presented showing what is possible to improve by simulating the climate at such resolution spanning from the better representation of spatial patterns and variability of precipitation at daily and hourly time scales, to the improved the representation of hourly frequency and intensity and summer diurnal cycle, representation of extreme events and reduction of model uncertainty.
The main community effort projects like EUCP and CORDEX-FPS Convection have been introduced and the main findings of the first ensemble of convection permitting climate simulations illustrated for both the present day validation and the climate projection. A summary of the key finding of similar extra European CP projects was also reported and discussed.

PASCAL YIOU – “SIMULATION OF RARE CLIMATE EVENTS”

Pascal Yiou made a lecture on the statistical methods to simulate extreme events. The goal was to provide the necessary information so that the students to facilitate a “hands on” experience in the afternoon.
The general mathematical background was defined and illustrated with examples. Care was taken to expose the mathematical formulation of the scientific questions and the strategies to treat them in a pedestrian fashion. This lead to general definitions of “rare event algorithms”.
A brief introduction to stochastic weather generators (SWGs) was proposed. SWGs are convenient tools to simulate very large ensembles of trajectories. The lecture was illustrated by a case study around the heatwave that struck north western US and Canada in the summer 2021. This illustrated how climate change affects the features of the most extreme events.

GABRIELE MESSORI – “LINKING LARGE-SCALE DYNAMICS TO REGIONAL EXTREMES”

Extreme climate events have multifarious detrimental impacts on both society at large and natural systems. The impacts of the extreme events themselves are often on a local to regional scale, but the physical processes that originate them can occur on much larger scales, up to several tens of thousands of kilometers. In this presentation, I review some of the large-scale atmospheric drivers of regional extremes, focusing on elements such as blocking, Rossby waves and jet streams. I also touch upon the role of these large-scale drivers in synchronizing extreme events across geographically remote regions.

THE TRAINING SESSIONS

HACKATHON 1 (20 JUNE) – “STATISTICAL METHODS FOR EXTREME EVENT ANALYSIS: HANDS ON” (DAVIDE FARANDA & AGLAÉ JÉZÉQUEL)

In the afternoon Aglaé Jézéquel and Davide Faranda performed an attribution exercise: the exercise consisted first on explanations on the statistical basis to perform extreme event attribution based on extreme value theory, then on a hands-on phase where the students could perform the attribution of extreme events using pre-prepared data. One of the most important outcomes of this exercise session was the fact that three of the students, namely Cadiou, Noyelle and Malhomme brought to publication their exercise (https://doi.org/10.1007/s13143-022-00305-1)

HACKATHON 2 (21 JUNE)- “DEEP LEARNING FOR SPATIO-TEMPORAL DROUGHT FORECASTING” (M.-Á. FERNÁNDEZ-TORRES)

In this hackathon, the students got a brief introduction to the use of PyTorch [Paszke et al., 2019] for the development of deep learning-based architectures [Goodfellow et al., 2016]. To that end, they had to implement and evaluate a multilayer perceptron for drought forecasting as a multi-class classification task in the United States [Minixhofer, C. et al., 2021].
The hackathon was conducted in a tutorial fashion using a Jupyter notebook, leaving time for the students to try to solve the different activities proposed. First, the database, its loading, normalization, and visualisation were presented. Then, the neural network was implemented, together with the loop for its execution during the training stage. Moreover, early stopping, which is a core strategy to select the best model in this phase, was reviewed. Last but not least, the confusion matrix was explained as an evaluation tool during the test stage of the system. Given the long duration of the hackathon, activities regarding early stopping and the confusion matrix were finished during Hackathon 3 session, which was directly related to it.
Students who chose the hackathon as a final project worked on the improvement of the training stage of the developed model and its extension on the basis of long short-term memory units, being able to outperform the initial baseline model.

HACKATHON 3 (22 JUNE) – “PRECIPITATION DOWNSCALING USING DEEP NEURAL NETWORKS” (M.-Á. FERNÁNDEZ-TORRES)

In this hackathon, the students developed a convolutional neural network (CNN) architecture [Goodfellow et al., 2016] for precipitation downscaling [Wang, J. et al., 2021] using PyTorch [Paszke et al., 2019]. The hackathon was conducted in a tutorial fashion using a Jupyter notebook, leaving time for the students to try to solve the different activities proposed. Given that the structure of the hackathon was similar to the one followed in Hackathon 2, its remaining activities (early stopping, confusion matrix) were solved first. Then, a baseline CNN architecture for precipitation downscaling was implemented, following the same procedure as in Hackathon 2 for training and evaluation purposes.
Students who chose the hackathon as a final project worked on the improvement of the developed model by changing its complexity (number of layers and filters per layer) and attempting to incorporate attention mechanisms.

CREATIVE WRITING WORKSHOP (23 JUNE) – (AGLAÉ JÉZÉQUEL)

Participants were asked to work in pairs. They started by a “silence bal” where they were supposed to only communicate with their partner in written form. They were then asked to think about what had fictional potential in their research, and to come up with pitches of short stories they could write about it. Some of them wrote short stories after the workshop. A network of potential writers was also established but I have lacked time to animate since the end of the school. There’s still potential for this network to evolve, and maybe to collect stories from the participants to the training school in the coming months.

OPEN DISCUSSION ABOUT HUMANITARIAN WORK, INTERESTS, NEEDS (27 JUNE) – (RCCC TEAM)

At the Summer School, we virtually facilitated an interactive session that covered topics that included the communication of attribution findings, the vulnerability and exposure context, and the potential impact of attribution studies on decision-making. We reflected that despite not being able to attend in person, we found that students were extremely interested in learning about our experience as a humanitarian organization working with attribution scientists and potentially using the findings. We engaged the students in a game and used the game mechanism to simulate consequences and rewards for making decisions based on attribution findings. The students asked thoughtful questions about the utility of different findings and discussed how they could be complemented or contextualized for different end-users. Overall, we found it a valuable experience to engage with a new generation of attribution scientists on critical questions that they will need to engage with as they do more of this work.

HACKATHON 4 (28 JUNE) – “ADVANCED METHODS FOR ATTRIBUTION, INCLUDING VULNERABILITY AND ADAPTATION ASSESSMENT” (ROBERT VAUTARD)

The activity of Hackathon 4 was to carry out in practice a WWA-type of attribution, applied to the case of the ongoing June 2022 heat wave. The exercise, made collectively by the group of students, led to a published XAIDA brief on the heat wave and is available on the XAIDA website: https://xaida.eu/early-2022-heat-waves-in-europe/

The exercise was organized in about 10 groups, each collecting data and exploring numbers by using the Climate Explorer tool https://climexp.knmi.nl developed by the KNMI partner of XAIDA. Each group was assigned to a subregion of Europe and studied how exceptional the event was and what was the change in statistical properties due to climate change.

HACKATHON 5 (29 JUNE) – “CONVECTION PERMITTING CLIMATE MODELLING DATA ANALYSIS” (EMANUELA PICCHELLI)

The activity of Hackathon 8 dedicated to atmospheric convection had the purpose to give the students an overview on how this phenomenon can be represented through regional climate models (RCM).
Conceptual aspects about atmospheric convection and its representation within the RCMs have been sketched, giving an overview on the advantages of simulating at the km scale for the reduction of model uncertainty and the proper representation of heavy precipitation events (HPE).
Then an exercise session followed. Observational and model data were available to the student and couple of exercise were proposed:
1. Analysis of extreme precipitation events case studies based on observations and assessment of ERA-Interim reanalysis products in reproducing such kind of events, whose purpose was to make the student aware about the concept of HPE and how this can or can not be represented in an area limited model.
2. Extended analysis over a statistical sample of extreme precipitation events by means of an appropriate statistical procedure. The exercise was shaped to make the students to:

a. Build a statistical analysis of precipitation. Example scripts were available, but also self-built analysis was welcome. A 10-years ICTP model dataset of precipitation was available for the scope.
b. Select the most appropriate statistical indices among several for the analysis of HPEs within observations and RCMs at different resolutions and eventually draw added value of models at the km scale. Ready statistical data were available for framing a multi-model analysis in comparison with multi-
observational dataset.

c. Select the most appropriate methodology to analyse the model spread/uncertainty for the above selected quantities. Two to three student groups were mainly interested in this activity and successfully performed the proposed analysis. Some of them were able to give a final presentation of their results.

HACKATHON 6 (30 JUNE) – “SIMULATING THE MOST EXTREME EVENTS” (PASCAL YIOU)

Pascal Yiou prepared computer codes and data sets on github, so that the students could reproduce the results presented in the lecture (made in the morning) and experiment by themselves on other regions (e.g. France, Europe), to simulate either hot or cold events. The students were able to download the prepared material from github and follow the training on the data sets. Two students (Camille Cadiou and Robin Noyelle) helped their fellows in bridging the R format into python (as most attendants were more familiar with python than R). In the end, one group chose to work on the hackathon of simulating extreme heatwave in the Iberian Peninsula. The results were promising, and were harbingers of the heatwave to come in that region during the summer.
This lecture/hackathon was the last of the school. The attendants hence had little time to work on that topic. I was immensely satisfied that a group of students was able to master the concepts and tools that were exposed to produce an interesting preliminary result in just a few hours.