Workshop on Uncertainty Quantification for Climate Science


When & where

The workshop took place on 12 and 13 November 2025 at Institut Henri Poincaré, Paris.

Room: amphithéâtre Hermite (ground floor).




Presentation of the workshop

This workshop was jointly organized by RT UQ, institut des Mathématiques pour la Planète Terre and GdR Défis theoriques pour les sciences du climat.

The objective of this workshop was to bring together members of different communities to discuss the methodological and practical aspects of uncertainty quantification in climate data, models and simulations. The topics covered included, but were not limited to: uncertainty quantification methodology, data assimilation, statistical modeling of extreme events, characterization and modeling of uncertainties in observational data, and the integration of machine learning tools into climate simulations.

The workshop consisted of two mini-courses on the afternoon of Wednesday 12 November, followed by a full day of specialized presentations on Thursday 13 November.

Organizers: Aurélie Fischer, Corentin Herbert, Clémentine Prieur, Julien Reygner, Jean-Noël Thépaut.


Schedule and slides

Wednesday 12 November (minicourses)

13:15 - 13:30 Welcome
13:30 - 15:30 Eric Blayo: An introduction to data assimilation (slides)
15:30 - 16:00 Coffee break
16:00 - 18:00 Josselin Garnier: An introduction to uncertainty quantification (slides)

Thursday 13 November (research talks)

9:20 - 9:30 Welcome
9:30 - 10:15 Pierre Tandeo: Quantifying uncertainty in climate reanalyses (slides)
10:15 - 10:45 Coffee break
10:45 - 11:30 Marc Bocquet: Stochastic surrogate models, data assimilation, and uncertainty quantification (slides)
11:30 - 12:15 Maxime Taillardat: Correcting and using uncertainty quantification through statistical post-processing (slides)
12:15 - 14:00 Lunch (buffet at IHP)
14:00 - 14:45 Claire Monteleoni: Generative AI for Climate and Weather
14:45 - 15:30 Chris Merchant: Satellite Observations of Climate Change: Uncertainty and Stability (slides)
15:30 - 16:00 Coffee break
16:00 - 16:45 Laure Raynaud: UQ in weather forecasting: a practical view for operational applications, and the new paradigm of machine learning (slides)
16:45 - 17:30 Laurent Dubus: Uncertainty Matters: Implications for Energy Planning and Power System Resilience (slides)


Mini-courses (Wednesday 12 November afternoon)

This mini-course aims to provide a simple introduction to data assimilation for non-specialists. It will cover the objectives, applications, and main families of methods of data assimilation, as well as the common challenges associated with it.

This mini-course is an introduction to uncertainty quantification for non-specialists: fundamentals of probabilistic modeling, sources of uncertainty, forward and inverse problems, propagation of uncertainty, surrogate modeling and emulation, sensitivity analysis, Bayesian inference and connection to data assimilation, decision-making under uncertainty.


Specialized talks (Thursday 13 November)

Recent advances in deep learning have enabled the construction of fast surrogate models for complex geophysical systems. Through their adjoints, these models also make it possible to carry out variational data assimilation. I will illustrate these developments using both deterministic and stochastic surrogates trained on a state-of-the-art physical sea-ice model. The resulting surrogates are stable, accurate, and physically consistent, and can be integrated into an operational sea-ice data assimilation and forecasting system.

In a second example, I will show that sequential data assimilation itself can be learned, yielding methods that are significantly more efficient and robust than current state-of-the-art approaches. This, in turn, opens new algorithmic directions for the theory of data assimilation.

In both cases, the success of the neural network models critically relies on their ability to efficiently represent and quantify uncertainty.

Uncertainty is an inherent feature of climate data, models, and simulations — and its influence extends far beyond the scientific domain. In energy planning, particularly in resource adequacy assessments and power system resilience studies, the treatment of uncertainty plays a critical role in shaping decisions and strategies. This presentation offers a general overview of how uncertainties—stemming from climate projections, demand forecasts, and infrastructure vulnerabilities—impact the reliability and robustness of energy systems.
By highlighting challenges and practical implications, the talk aims to foster dialogue between climate scientists, energy modelers, and decision-makers, and to advocate for integrated approaches that better account for uncertainty in long-term energy planning.

Observations from space-based sensors provide global insights into the changes on Earth across a range of essential climate variables. The climate data records (CDRs) derived from these satellite observations have varied and complex sources of error, which makes quantifying their uncertainty a challenge. Understanding both the magnitude of observational uncertainty and the error correlation properties is essential for correctly inferring the uncertainty in derived quantities, such as temporal differences, climate indices, modes of variability, and climate trends.

A framework for organising and focusing the evaluation of uncertainty can be based on metrological principles adapted to the specific circumstances of Earth observations. While the uncertainty characterisation of CDRs in this manner requires significant effort, the process also challenges and improves the ideas and methods used to create these records.
A particular problem arises when quantifying long-term change. In the climate literature, the uncertainty in temporal trends is often discussed without considering observational stability. This is unsurprising, given that CDR creators usually don't provide information about stability. Furthermore, the discussion of stability in the context of requirements-setting for essential climate variables has arguably hindered progress in assessing it. Proposals to clarify the concept of stability will help creators and users of CDRs make more realistic evaluations of uncertainty in climate trends.

The stunning recent advances in AI content generation rely on cutting-edge, generative deep learning algorithms and architectures trained on massive amounts of text, image, and video data. With different training data, these algorithms and architectures can benefit a variety of applications for addressing climate change. As opposed to text and video, the relevant training data includes weather and climate data from observations, reanalyses, and even physical simulations.

Using AI methods — especially generative AI — in climate science provides additional sources of uncertainty beyond those already identified. However the potential for such methods is great. Our team has shown the benefit of generative AI methods for data fusion, interpolation, downscaling, and domain alignment. I will provide a survey of our recent work applying these methods to problems including weather forecasting, with a particular focus on extreme events, climate model emulation and scenario generation, and renewable energy planning.

Quantifying uncertainty is of utmost importance for weather forecasting due to the chaotic nature of the atmosphere. This has led to the development of Monte Carlo methods under the form of ensemble prediction systems (EPSs), that are now key in a number of operational weather services worldwide. In the first part of the presentation we will review the state-of-the-art and some well-established methods to account for the main sources of uncertainty, regarding initial conditions and model formulation. The current view of UQ is, however, facing new challenges with the next-generation of forecasting systems that target higher resolution (100-m scale) and are likely to rely on the new paradigm of data-driven modeling to a large extent. The second part of the presentation will be dedicated to UQ and machine learning. We will present early works conducted at Météo-France to leverage generative machine learning to improve UQ and in particular the sampling of operational EPSs. We will finally discuss how UQ could be adressed in data-driven systems and some open questions.

In weather forecasting, ensemble forecasts seek to represent and quantify different uncertainty sources in the forecast such as observation errors or a mathematical representation of the atmosphere still incomplete or based on approximations. They tend to be biased and misdispersed, affecting their reliability and resolution. Statistical post-processing, from simple bias corrections to the use of the most recent AI/ML techniques, is now a key component of the forecasting suites. In a first part we will show how such a correction of ensemble forecasts is made with examples coming from the operational forecast chain of Météo-France. In a second part, we will adress a question some users can have: how to summarize optimally the information of the uncertainty quantification into a single outcome? With a theoretical proof and an example over precipitation in France, we will present a new type of point forecasts, called crossing-point quantile, better suited for forecasting some events than existing approaches, at least for some users.

Climate reanalyses such as ERA5 allow the study of numerous variables over long periods. These reanalyses integrate a large number of historical observations. These observations are assimilated into a general circulation model. Most of the time, reanalyses data are obtained using ensemble filtering methods that allow for the estimation of uncertainties in the reanalyzed fields. These assimilation uncertainties, also known as a posteriori errors, depend on two factors: the model error covariance, denoted Q and referring to the a priori error, and the observation error covariance, denoted R, which takes into account instrumentation error and representativeness error.

In this talk, I will demonstrate the importance of the Q and R covariances in estimating the uncertainty of data assimilation algorithms. I will review several algorithms for jointly estimating the Q and R matrices. I will also introduce a new metric to quantify the uncertainties of reanalyses: the credibility score. It is based on the notion of confidence interval, coverage probability, and allows us to check whether uncertainties are underestimated or overestimated.


List of participants