2025Activity​​​‌ reportProject-TeamARCHES

2.1.1‌​‌ Research objectives

ARCHES: AI​​ Research for Climate Change​​​‌ and Environmental Sustainability is‌ focused on using AI‌​‌ to address climate change,​​ and to enable environmentally​​​‌ sustainable solutions. In particular,‌ ARCHES research is focused‌​‌ along three axes:

1. AI​​ for Climate Change Adaptation​​​‌ – Forecasting and informing‌ near-term decisions
2. AI for‌​‌ Climate Change Mitigation –​​ Forecasting and informing mid-term​​​‌ decisions
3. AI for Understanding‌ Climate Change Impacts –‌​‌ Projecting long-term impacts

While​​ addressing these problems, we​​​‌ will also continue to‌ advance core AI research‌​‌ in machine learning in​​ computer vision. Our past​​​‌ work has demonstrated that‌ climate and environmental applications‌​‌ open new questions for​​ the design and analysis​​​‌ of machine learning algorithms.‌ We have also found‌​‌ that applied research can​​ yield unorthodox twists, even​​​‌ on standard machine learning‌ techniques, which in turn‌​‌ spark interest in the​​ machine learning research community.​​​‌ Examples include our work‌ on climate prediction via‌​‌ sparse matrix completion, unsupervised​​ learning of data-driven, probabilistic​​​‌ definitions of diverse, multivariate‌ extreme events, and deep‌​‌ unsupervised learning for anomaly​​ detection in problems with​​​‌ limited labeled data.

7.1.1 DRAGON

• Name:
  Directed​ Acyclic Graph OptimisatiON
• Keywords:​‌
  Automated machine learning, Neural​​ architecture search, Neural networks​​
• Scientific Description:
  The search​​​‌ space is made from‌ Python objects called Variables,‌​‌ which can encode integers,​​ arrays, or other elements.​​​‌ These variables are combined‌ to create DAGs representing‌​‌ neural networks. Each variable​​ can also have neighbor​​​‌ or mutation operators to‌ explore slightly different configurations,‌​‌ and crossover operators allow​​ combining multiple configurations. DRAGON​​​‌ includes several search algorithms:‌ Random Search, Evolutionary Algorithm,‌​‌ Mutant-UCB, and HyperBand. Some​​ algorithms (like Evolutionary Algorithm​​​‌ and Mutant-UCB) use the‌ neighbor/mutation functions to explore‌​‌ new configurations. Algorithms also​​ come with memory-efficient storage​​​‌ and an optional distributed‌ version for running on‌​‌ multiple processors. Finally, performance​​ evaluation is handled by​​​‌ the user: a network‌ is built from a‌​‌ configuration, trained, and evaluated​​ to return a loss,​​​‌ which the search algorithms‌ aim to minimize.
• Functional‌​‌ Description:
  DRAGON is an​​ open-source Python tool that​​​‌ helps design and optimize‌ deep learning models. It‌​‌ represents neural networks as​​ graphs of connected operations,​​​‌ where each operation can‌ be adjusted to improve‌​‌ performance. Unlike many “automatic”​​ AI tools, DRAGON requires​​​‌ users to define the‌ structure of the network‌​‌ and how it is​​ trained, giving much more​​​‌ flexibility to solve different‌ problems. Users can explore‌​‌ different network designs by​​ slightly modifying existing configurations​​​‌ or combining multiple designs.‌ The tool includes methods‌​‌ to test many options​​ efficiently, keeping memory usage​​​‌ low and even allowing‌ multiple computers to work‌​‌ together. Once a network​​ is designed, it is​​​‌ trained and evaluated, and‌ the results guide the‌​‌ next round of improvements.​​ DRAGON has been utilized​​​‌ for tasks such as‌ image recognition and predicting‌​‌ energy load, but it​​ can be adapted to​​​‌ address many other AI‌ problems.
• URL:
  https://­github.­com/­JulieKeisler/­DRAGON
• Contact:‌​‌
  Julie Keisler

7.1.2 geoarches​​

• Keywords:
  Machine learning, Forecasting,​​​‌ Climate change, Data processing‌
• Scientific Description:
  geoarches is‌​‌ a research-friendly machine learning​​ library for training, running,​​​‌ and evaluating models on‌ geospatial data, mainly weather‌​‌ and climate data. Built​​ on PyTorch, Pytorch Lightning,​​​‌ and Hydra, geoarches offers‌ a clean, modular structure‌​‌ for developing and scaling​​ ML pipelines. It can​​​‌ also be used to‌ run the ArchesWeather and‌​‌ ArchesWeatherGen weather models.
• Functional​​ Description:
  geoarches is a​​​‌ machine learning library for‌ training, running and evaluating‌​‌ models on weather and​​ climate data.
• Release Contributions:​​​‌
  This is the first‌ version.
• URL:
  https://­github.­com/­INRIA/­geoarches
• Contact:‌​‌
  Renu Singh

7.1.3 SerpentFlow​​

• Name:
  SharEd-structuRe decomPosition for​​​‌ gEnerative domaiN adapTation
• Keywords:‌
  Deep learning, Super-resolution, Domain‌​‌ Adaptation
• Functional Description:
  SerpentFlow​​ (SharEd-structuRe decomPosition for gEnerative​​​‌ domaiN adapTation) is a‌ framework for unpaired domain‌​‌ alignment. It separates shared​​ low-frequency structures from domain-specific​​​‌ high-frequency content and uses‌ Flow Matching for generative‌​‌ modeling.
• Contact:
  Julie Keisler​​

7.1.4 Motif

• Name:
  Multi-source​​​‌ transformer via factorized attention‌
• Keywords:
  Forecasting, Deep learning,‌​‌ Climate change, Satellite imagery​​
• Scientific Description:

  This repository​​​‌ implements a DL architecture‌ adapted to learning from‌​‌ multiple sources. The possibilities​​ of inputs include:

  A​​​‌ flexible number of sources:‌ while a large set‌​‌ of sources can be​​ used as input to​​​‌ the model, a specific‌ sample may contain any‌​‌ subset of the sources.​​​‌ Samples with different numbers​ of sources can be​‌ bathed together during training​​ or inference. Sources of​​​‌ different dimensionalities and natures​ (0D, e.g. station measurements,​‌ 1D, e.g. vertical profile,​​ 2D, e.g. remote sensing​​​‌ images). Sources misaligned in​ space and time, for​‌ example remote sensing images​​ covering different geographical areas​​​‌ (which may even be​ disjoint), and at irregular​‌ time intervals. Sources of​​ the same type with​​​‌ different characteristics, e.g. remote​ sensing images in the​‌ same band from different​​ satellites with different exact​​​‌ frequencies and ground sampling​ distance

• Functional Description:
  Motif​‌ is a Python package​​ to train AI models​​​‌ on geospatial data from​ multiple sources. The code​‌ includes a data engineering​​ pipeline and a novel​​​‌ neural networks architecture for​ data fusion.
• Contact:
  Clement​‌ Dauvilliers

8.1.1 Climate change dependence​​ on local time

Participants:​​​‌ Sarah Safieddine.

In​ 14, we calculate​‌ climate trends in local​​ time. In fact, essential​​​‌ Climate Variables, such as​ near-surface (T2m) and land​‌ surface temperatures (LST), are​​ typically reported in Coordinated​​​‌ Universal Time (UTC) for​ global consistency. However, their​‌ diurnal variability leads to​​ temperature trends that differ​​​‌ by the local hour,​ a factor not analyzed​‌ on the global nor​​ regional scale. Using ECMWF​​​‌ ERA5-Land reanalysis data (1981–2022),​ we assess temperature trends​‌ by local hour and​​ month. Our results show​​​‌ that the trends can​ change significantly during the​‌ day. LST and T2m​​ warming or cooling trends​​​‌ peak in the afternoon,​ while showing large spatial​‌ variability across both hemispheres.​​ Using MODIS observations, we​​​‌ show how the nominal​ Equator crossing times of​‌ TERRA and AQUA influence​​ LST trends. These findings​​​‌ highlight the necessity of​ accounting for local time​‌ in climate assessments to​​ improve adaptation strategies.

8.1.2​​​‌ AI for weather forecasting​

Participants: Guillaume Couairon,​‌ Renu Singh, Anastase​​ Charantonis, Claire Monteleoni​​​‌.

Weather forecasting plays​ a vital role in​‌ today's society, from agriculture​​ and logistics to predicting​​​‌ the output of renewable​ energies, and preparing for​‌ extreme weather events. Deep​​ learning weather forecasting models​​​‌ trained with the next​ state prediction objective on​‌ ERA5 have shown great​​ success compared to numerical​​​‌ global circulation models. However,​ for a wide range​‌ of applications, being able​​ to provide representative samples​​​‌ from the distribution of​ possible future weather states​‌ is critical. In 1​​, we propose a​​​‌ methodology to leverage deterministic​ weather models in the​‌ design of probabilistic weather​​ models, leading to improved​​​‌ performance and reduced computing​ costs. We first introduce​‌ ArchesWeather, a transformer-based deterministic​​ model that improves upon​​​‌ Pangu-Weather by removing overrestrictive​ inductive priors. We then​‌ design a probabilistic weather​​ model called ArchesWeatherGen based​​ on flow matching, a​​​‌ modern variant of diffusion‌ models, that is trained‌​‌ to project ArchesWeather's predictions​​ to the distribution of​​​‌ ERA5 weather states. ArchesWeatherGen‌ is a true stochastic‌​‌ emulator of ERA5 and​​ surpasses IFS ENS and​​​‌ NeuralGCM on all WeatherBench‌ headline variables (except for‌​‌ NeuralGCM's geopotential). Our work​​ also aims to democratize​​​‌ the use of deterministic‌ and generative machine learning‌​‌ models in weather forecasting​​ research, with academic computing​​​‌ resources. All models are‌ trained at 1.5° resolution,‌​‌ with a training budget​​ of approximately 9 V100​​​‌ days for ArchesWeather and‌ 45 V100 days for‌​‌ ArchesWeatherGen. For inference, ArchesWeatherGen​​ generates 15-day weather trajectories​​​‌ at a rate of‌ 1 minute per ensemble‌​‌ member on a A100​​ GPU card. To make​​​‌ our work fully reproducible,‌ our code and models‌​‌ are open source, including​​ the complete pipeline for​​​‌ data preparation, training, and‌ evaluation, accessible here.‌​‌

Participants: David Landry,​​ Claire Monteleoni, Anastase​​​‌ Charantonis.

In 11‌, we propose a‌​‌ machine‐learning‐based methodology for in​​ situ weather forecast postprocessing​​​‌ that is both spatially‌ coherent and multivariate. Compared‌​‌ with previous work, our​​ Flow MAtching Postprocessing (FMAP)​​​‌ represents the correlation structures‌ of the observation distribution‌​‌ better, while also improving​​ marginal performance at stations.​​​‌ FMAP generates forecasts that‌ are not bound to‌​‌ what is already modeled​​ by the underlying gridded​​​‌ prediction and can infer‌ new correlation structures from‌​‌ data. The resulting model​​ can generate an arbitrary​​​‌ number of forecasts from‌ a limited number of‌​‌ numerical simulations, allowing for​​ low‐cost forecasting systems. A​​​‌ single training is sufficient‌ to perform postprocessing at‌​‌ multiple lead times, in​​ contrast with other methods,​​​‌ which use multiple trained‌ networks at generation time.‌​‌ This work details our​​ methodology, including a spatial​​​‌ attention transformer backbone trained‌ within a flow‐matching generative‌​‌ modeling framework. FMAP shows​​ promising performance in experiments​​​‌ on the EUMETNET Postprocessing‌ Benchmark (EUPPBench ) dataset,‌​‌ forecasting surface temperature and​​ wind‐gust values at station​​​‌ locations in western Europe‌ up to five‐day lead‌​‌ times.

Participants: Cecile Mallet​​.

A major issue​​​‌ limiting the successful deployment‌ of deep learning algorithms‌​‌ in geophysical applications is​​ their inability to generalize​​​‌ to new contexts. Regarding‌ the quantitative precipitation estimation‌​‌ (QPE) from the Global​​ Precipitation Mission (GPM) satellite​​​‌ constellation, the GPM Microwave‌ Imager (GMI) contains enough‌​‌ co-located brightness temperatures and​​ rain rates data to​​​‌ train a deep learning‌ inverse model to retrieve‌​‌ precipitation intensity. However, the​​ difference in instrumental configurations​​​‌ makes it impossible to‌ directly apply this inverse‌​‌ operator to another space-borne​​ radiometric imager. A domain​​​‌ adaptation is thus necessary‌ to solve the domain‌​‌ shift problem encountered when​​ applying the model trained​​​‌ on one satellite to‌ another satellite. The paper,‌​‌ 16, tests a​​ method to map the​​​‌ SSMI/S data to the‌ GMI data. In the‌​‌ absence of sufficient paired​​ images between the two​​​‌ satellites, we applied a‌ Cycle consistent Generative Adversarial‌​‌ Network (CycleGAN), which allows​​ for an Unsupervised Domain​​​‌ Adaptation approach. Evaluating the‌ quality of adapted images‌​‌ is a complex problem.​​​‌ This paper employs two​ tactics: a brief evaluation​‌ of adapted radiometric images​​ and a qualitative/quantitative evaluation​​​‌ of rain retrieval. Over​ several case studies, the​‌ results show that the​​ domain adaptation step produces​​​‌ adapted SSMI/S images that​ retain the majority of​‌ the rain structure. Next,​​ the rain detection score​​​‌ and intensity bias are​ then compared using 847​‌ overpasses. The same analysis​​ is carried out over​​​‌ mainland France by comparing​ the results with rainfall​‌ products supplied by Météo-France.​​ In both comparisons, the​​​‌ adapted images allow the​ inverse operator to provide​‌ a better score in​​ rain detection and intensity.​​​‌

8.1.3 Subseasonal wind forecasting​

Participants: Ganglin Tian,​‌ Anastase Alexandre Charantonis.​​

In 17 to improve​​​‌ the spatial representation of​ uncertainties when regressing surface​‌ wind speeds from large-scale​​ atmospheric predictors for sub-seasonal​​​‌ forecasting. Sub-seasonal forecasting often​ relies on large-scale atmospheric​‌ predictors such as 500​​ hPa geopotential height (Z500),​​​‌ which exhibit higher predictability​ than surface variables and​‌ can be downscaled to​​ obtain more localised information.​​​‌ Previous work by Tian​ et al. (2024) demonstrated​‌ that stochastic perturbations based​​ on model residuals can​​​‌ improve ensemble dispersion representation​ in statistical downscaling frameworks,​‌ but this method fails​​ to represent spatial correlations​​​‌ and physical consistency adequately.​ More sophisticated approaches are​‌ needed to capture the​​ complex relationships between large-scale​​​‌ predictors and local-scale predictands​ while maintaining physical consistency.​‌ Probabilistic deep learning models​​ offer promising solutions for​​​‌ capturing complex spatial dependencies.​ This study evaluates three​‌ probabilistic methods with distinct​​ uncertainty quantification mechanisms: Quantile​​​‌ Regression Neural Network that​ directly models distribution quantiles,​‌ Variational Autoencoders that leverage​​ latent space sampling, and​​​‌ Diffusion Models that utilise​ iterative denoising. These models​‌ are trained on ERA5​​ reanalysis data and applied​​​‌ to ECMWF sub-seasonal hindcasts​ to regress probabilistic wind​‌ speed ensembles. Our results​​ show that probabilistic downscaling​​​‌ approaches provide more realistic​ spatial uncertainty representations compared​‌ to simpler stochastic methods,​​ with each probabilistic model​​​‌ offering different strengths in​ terms of ensemble dispersion,​‌ deterministic skill, and physical​​ consistency. These findings establish​​​‌ probabilistic downscaling as an​ effective enhancement to operational​‌ sub-seasonal wind forecasts for​​ renewable energy planning and​​​‌ risk assessment.

8.1.4 Oceanic​ data interpolation

Participants: Dmitrii​‌ Drozdov, Pierre Garcia​​, Anastase Alexandre Charantonis​​​‌.

In 4,​ we propose a novel​‌ method for reconstruction of​​ high-resolution Sea Surface Height​​​‌ (SSH) fields from sparse​ along-track satellite altimetry. We​‌ explore the usage of​​ an observation-driven deep-learning method​​​‌ for inpainting, with a​ focus on diffusion-based generative​‌ models for spatial reconstruction​​ and data assimilation. The​​​‌ study includes data preparation​ from reanalyses and satellite​‌ observations, the definition of​​ evaluation metrics relevant to​​​‌ oceanography, and comparison with​ baselines. We discuss model​‌ design choices, uncertainty characterization​​ through stochastic sampling, and​​​‌ limitations in real-world deployment.​ Effectively, in this work,​‌ we develop and validate​​ an observation-driven prior, allowing​​​‌ us to sample from​ the ground-truth distribution of​‌ SSH. By not relying​​ on simulation results for​​​‌ training, we propose a​ step towards observationdriven Deep-Learning​‌ analysis of SSH and​​ its uncertainties at small​​ scales

8.1.5 Oceanic data​​​‌ forecasting and assimilation

Participants:‌ Anastase Alexandre Charantonis.‌​‌

Abstract Sea Surface Height​​ Anomaly (SLA) is a​​​‌ signature of the mesoscale‌ dynamics of the upper‌​‌ ocean. Sea surface temperature​​ (SST) is driven by​​​‌ these dynamics and can‌ be used to improve‌​‌ the spatial interpolation of​​ SLA fields. In 13​​​‌ we focused on the‌ temporal evolution of SLA‌​‌ fields. We explored the​​ capacity of deep learning​​​‌ (DL) methods to predict‌ short-term SLA fields using‌​‌ SST fields. We used​​ simulated daily SLA and​​​‌ SST data from the‌ Mercator Global Analysis and‌​‌ Forecasting System, with a​​ resolution of (1/12)° in​​​‌ the North Atlantic Ocean‌ (26.5-44.42°N, -64.25-41.83°E), covering the‌​‌ period from 1993 to​​ 2019. Using a slightly​​​‌ modified image-to-image convolutional DL‌ architecture, we demonstrated that‌​‌ SST is a relevant​​ variable for controlling the​​​‌ SLA prediction. With a‌ learning process inspired by‌​‌ the teaching-forcing method, we​​ managed to improve the​​​‌ SLA forecast at 5‌ days by using the‌​‌ SST fields as additional​​ information. We obtained predictions​​​‌ of 12 cm (20‌ cm) error of SLA‌​‌ evolution for scales smaller​​ than mesoscales and at​​​‌ time scales of 5‌ days (20 days) respectively.‌​‌ Moreover, the information provided​​ by the SST allows​​​‌ us to limit the‌ SLA error to 16‌​‌ cm at 20 days​​ when learning the trajectory.​​​‌

Participants: Pierre Garcia,‌ Anastase Alexandre Charantonis.‌​‌

In 5, we​​ explore the capacity of​​​‌ recent diffusion and flow‌ matching techniques for data‌​‌ assimilation of oceanic, sparsely​​ observed fields. Providing regular​​​‌ and physically consistent predictions‌ of the ocean state‌​‌ is critical for numerous​​ scientific, operational, and societal​​​‌ needs. Observations of the‌ ocean surface are gathered‌​‌ through various remote sensing​​ and in situ instruments,​​​‌ and are typically assimilated‌ into numerical models to‌​‌ reconstruct the ocean state.​​ However, this often involves​​​‌ millions of data points,‌ making it computationally intensive,‌​‌ which suggests deep learning​​ may be a cheaper​​​‌ alternative. Deterministic data-driven approaches‌ typically learn about ocean‌​‌ dynamics from numerical simulations​​ or sparse observational data.​​​‌ However, such methods often‌ lack physical realism in‌​‌ uncertain settings. Due to​​ mode averaging, they produce​​​‌ non-physical or overly simplified‌ states. Generative models offer‌​‌ a promising approach to​​ generating physically realistic ocean​​​‌ states. We present GloFM:‌ a Glorys Flow-Matching emulator‌​‌ for spatio-temporal ocean data​​ assimilation. Our generative model​​​‌ produces coherent estimates of‌ ocean surface fields. GloFM‌​‌ uses flow matching to​​ assimilate observational data for​​​‌ nowcasting of surface currents,‌ sea surface height (SSH),‌​‌ and sea surface temperature​​ (SST). Compared to deterministic​​​‌ regression-based approaches, GloFM demonstrates‌ improved realism metrics, capturing‌​‌ finer-scale variability and more​​ physically plausible ocean states.​​​‌

8.2.1 New AI tools‌​‌ for energy forecasting

Participants:​​ Julie Keisler.

Current​​​‌ technologies only allow storage‌ by expensive and inefficient‌​‌ means, which makes it​​ difficult to store electricity​​​‌ on a large scale.‌ For the grid to‌​‌ function properly, electricity fed​​ into the grid must​​​‌ match electricity used at‌ all times. Historically, and‌​‌ still today, production resources​​​‌ are planned in advance​ of demand to maintain​‌ this balance. It is​​ therefore crucial to forecast​​​‌ electricity consumption as accurately​ as possible. The integration​‌ of renewable energies, whose​​ production is intermittent and​​​‌ dependent on weather conditions,​ is making the balance​‌ increasingly unstable. Managing this​​ is becoming more complex,​​​‌ making forecasting wind and​ photovoltaic production now essential.​‌ Statistical learning models are​​ used to make consumption​​​‌ and production forecasts. These​ models take past values​‌ and data from explanatory​​ variables and use them​​​‌ to model the signal.​ To build efficient models,​‌ one must choose the​​ input variables, the type​​​‌ of model, and its​ parameters. Given the vast​‌ number of signals to​​ be forecasted, it would​​​‌ be beneficial to automate​ these choices to create​‌ competitive models.

Automated Machine​​ Learning (AutoML) is the​​​‌ process of automating the​ generation of learning models​‌ optimized according to the​​ use case. Over the​​​‌ last ten years, numerous​ AutoML tools have been​‌ developed. However, most of​​ them focus on optimizing​​​‌ classification or regression models​ on tabular data, or​‌ on optimizing neural network​​ architectures for image or​​​‌ text processing. These tools​ are not appropriate for​‌ optimizing electricity consumption and​​ production forecasting models. This​​​‌ thesis is a progress​ towards automating the generation​‌ of time series forecasting​​ models required for power​​​‌ system management. 10 focused​ on developing the DRAGON​‌ Python package, which offers​​ a range of tools​​​‌ for specific yet widely​ used models: neural networks.​‌ DRAGON can be used​​ to create flexible search​​​‌ spaces encompassing a wide​ variety of neural networks​‌ by simultaneity optimizing the​​ architecture and the hyperparameters.​​​‌ They are encoded by​ Directed Acyclic Graphs (DAGs),​‌ where the nodes are​​ operations, parameterised by various​​​‌ hyperparameters, and the edges​ are the connections between​‌ these nodes. To navigate​​ these graph-based search spaces​​​‌ and optimize their structures,​ the package proposes various​‌ search algorithms based on​​ meta-heuristics and bandits-approaches. This​​​‌ thesis details how DRAGON​ is used for electricity​‌ consumption and production forecasts,​​ enabling state-of-the-art models to​​​‌ be generated for these​ two industrial use cases.​‌

Electricity demand forecasting is​​ key to ensuring that​​​‌ supply meets demand lest​ the grid would blackout.​‌ Reliable short-term forecasts may​​ be obtained by combining​​​‌ a Generalized Additive Models​ (GAM) with a State-Space​‌ model, leading to an​​ adaptive (or online) model.​​​‌ A GAM is an​ over-parameterized linear model defined​‌ by a formula and​​ a state-space model involves​​​‌ hyperparameters. Both the formula​ and adaptation parameters have​‌ to be fixed before​​ model training and have​​​‌ a huge impact on​ the model's predictive performance.​‌ In 2, we​​ propose optimizing them using​​​‌ the DRAGON package mentioned​ above, originally designed for​‌ neural architecture search. This​​ work generalizes it for​​​‌ automated online generalized additive​ model selection by defining​‌ an efficient modeling of​​ the search space (namely,​​​‌ the space of the​ GAM formulae and adaptation​‌ parameters). Its application to​​ short-term French electricity demand​​​‌ forecasting demonstrates the relevance​ of the approach

8.3.1 Climate model​​ emulation

Participants: Graham Clyne​​​‌, Guillaume Couairon,‌ Claire Monteleoni, Anastase‌​‌ Charantonis.

Climate projections​​ have uncertainties related to​​​‌ components of the climate‌ system and their interactions.‌​‌ A typical approach to​​ quantifying these uncertainties is​​​‌ to use climate models‌ to create ensembles of‌​‌ repeated simulations under different​​ initial conditions. Due to​​​‌ the complexity of these‌ simulations, generating such ensembles‌​‌ of projections is computationally​​ expensive. In 27,​​​‌ we present ArchesClimate, a‌ deep learning-based climate model‌​‌ emulator that aims to​​ reduce this cost. ArchesClimate​​​‌ is trained on decadal‌ hindcasts of the IPSL-CM6A-LR‌​‌ climate model at a​​ spatial resolution of approximately​​​‌ 2.5x1.25 degrees. We train‌ a flow matching model‌​‌ following ArchesWeatherGen, which we​​ adapt to predict near-term​​​‌ climate. Once trained, the‌ model generates states at‌​‌ a one-month lead time​​ and can be used​​​‌ to auto-regressively emulate climate‌ model simulations of any‌​‌ length. We show that​​ for up to 10​​​‌ years, these generations are‌ stable and physically consistent.‌​‌ We also show that​​ for several important climate​​​‌ variables, ArchesClimate generates simulations‌ that are interchangeable with‌​‌ the IPSL model. This​​ work suggests that climate​​​‌ model emulators could significantly‌ reduce the cost of‌​‌ climate model simulations.

8.4.1 A‌​‌ novel preconditioning-inspired iterative approach​​ to solving PDEs with​​​‌ neural networks

Participants: Emmanuel‌ de Bézenac.

In‌​‌ 12, physics-informed deep​​ learning often faces optimization​​​‌ challenges due to the‌ complexity of solving partial‌​‌ differential equations (PDEs), which​​ involve exploring large solution​​​‌ spaces, require numerous iterations,‌ and can lead to‌​‌ unstable training. These challenges​​ arise particularly from the​​​‌ ill-conditioning of the optimization‌ problem caused by the‌​‌ differential terms in the​​ loss function. To address​​​‌ these issues, we propose‌ learning a solver, i.e.,‌​‌ solving PDEs using a​​ physics-informed iterative algorithm trained​​​‌ on data. Our method‌ learns to condition a‌​‌ gradient descent algorithm that​​ automatically adapts to each​​​‌ PDE instance, significantly accelerating‌ and stabilizing the optimization‌​‌ process and enabling faster​​ convergence of physics-aware models.​​​‌ Furthermore, while traditional physics-informed‌ methods solve for a‌​‌ single PDE instance, our​​ approach extends to parametric​​​‌ PDEs. Specifically, we integrate‌ the physical loss gradient‌​‌ with PDE parameters, allowing​​ our method to solve​​​‌ over a distribution of‌ PDE parameters, including coefficients,‌​‌ initial conditions, and boundary​​ conditions. We demonstrate the​​​‌ effectiveness of our approach‌ through empirical experiments on‌​‌ multiple datasets, comparing both​​ training and test-time optimization​​​‌ performance. The code is‌ available at .

8.4.2‌​‌ A new model for​​ multi-source data fusion via​​​‌ self-supervised learning

Participants: Clément‌ Dauvilliers, Claire Monteleoni‌​‌.

In 3,​​ we present a deep​​​‌ learning architecture that reconstructs‌ a source of data‌​‌ at given spatio-temporal coordinates​​ using other sources. The​​​‌ model can be applied‌ to multiple sources in‌​‌ a broad sense: the​​ number of sources may​​​‌ vary between samples, the‌ sources can differ in‌​‌ dimensionality and sizes, and​​ cover distinct geographical areas​​​‌ at irregular time intervals.‌ The network takes as‌​‌ input a set of​​​‌ sources that each include​ values (e.g., the pixels​‌ for two-dimensional sources), spatio-temporal​​ coordinates, and source characteristics.​​​‌ The model is based​ on the Vision Transformer,​‌ but separately embeds the​​ values and coordinates and​​​‌ uses the embedded coordinates​ as relative positional embedding​‌ in the computation of​​ the attention. To limit​​​‌ the cost of computing​ the attention between many​‌ sources, we employ a​​ multi-source factorized attention mechanism,​​​‌ introducing an anchor-points-based cross-source​ attention block. We name​‌ the architecture MoTiF (multi-source​​ transformer via factorized attention).​​​‌ We present a self-supervised​ setting to train the​‌ network, in which one​​ source chosen randomly is​​​‌ masked and the model​ is tasked to reconstruct​‌ it from the other​​ sources. We test this​​​‌ self-supervised task on tropical​ cyclone (TC) remote-sensing images,​‌ ERA5 states, and best-track​​ data. We show that​​​‌ the model is able​ to perform TC ERA5​‌ fields and wind intensity​​ forecasting from multiple sources,​​​‌ and that using more​ sources leads to an​‌ improvement in forecasting accuracy.​​

8.4.3 A new generative​​​‌ domain alignment algorithm via​ shared-structure decomposition

Participants: Julie​‌ Keisler, Anastase Charantonis​​, Claire Monteleoni.​​​‌

Domain alignment refers broadly​ to learning correspondences between​‌ data distributions from distinct​​ domains. In this work,​​​‌ we focus on a​ setting where domains share​‌ underlying structural patterns despite​​ differences in their specific​​​‌ realizations. The task is​ particularly challenging in the​‌ absence of paired observations,​​ which removes direct supervision​​​‌ across domains. In 28​, we introduce a​‌ generative framework, called SerpentFlow​​ (SharEd-structuRe decomPosition for gEnerative​​​‌ domaiN adapTation), for unpaired​ domain alignment. SerpentFlow decomposes​‌ data within a latent​​ space into a shared​​​‌ component common to both​ domains and a domain-specific​‌ one. By isolating the​​ shared structure and replacing​​​‌ the domain-specific component with​ stochastic noise, we construct​‌ synthetic training pairs between​​ shared representations and target-domain​​​‌ samples, thereby enabling the​ use of conditional generative​‌ models that are traditionally​​ restricted to paired settings.​​​‌ We apply this approach​ to super-resolution tasks, where​‌ the shared component naturally​​ corresponds to low-frequency content​​​‌ while high-frequency details capture​ domain-specific variability. The cutoff​‌ frequency separating low- and​​ high-frequency components is determined​​​‌ automatically using a classifier-based​ criterion, ensuring a data-driven​‌ and domain-adaptive decomposition. By​​ generating pseudo-pairs that preserve​​​‌ low-frequency structures while injecting​ stochastic high-frequency realizations, we​‌ learn the conditional distribution​​ of the target domain​​​‌ given the shared representation.​ We implement SerpentFlow using​‌ Flow Matching as the​​ generative pipeline, although the​​​‌ framework is compatible with​ other conditional generative approaches.​‌ Experiments on synthetic images,​​ physical process simulations, and​​​‌ a climate downscaling task​ demonstrate that the method​‌ effectively reconstructs high-frequency structures​​ consistent with underlying low-frequency​​​‌ patterns, supporting shared-structure decomposition​ as an effective strategy​‌ for unpaired domain alignment.​​

10.1.1 Visits of​​​‌ international scientists

10.3.1 Invited talks/panels‌​‌ at EU events

Monteleoni:​​ European Central Bank Workshop/Conference​​​‌ on The Transformative Power‌ of AI: Economic Implications‌​‌ and Challenges, Frankfurt, Germany,​​ April 2025

Charantonis: “Securing​​​‌ European Digital Sovereignty: Evaluating‌ Global Dependencies, Risks, and‌​‌ the European Response,” Brussels,​​ October 2025

Monteleoni: EU​​​‌ Science for Preparedness Conference,‌ Turin, Italy, November 2025‌​‌

Monteleoni: "National Meteorological Services​​ and the EU :​​​‌ provide resilience in a‌ changing climate and fostering‌​‌ European innovation," Brussels, November​​ 2025

11.1.1 Scientific events: organisation‌

11.1.2​​​‌ Scientific events: selection

11.1.3 Journal

11.1.4 Invited talks

Selected​‌ Invited Talks

Emmanuel de​​ Bézenac: Scientific Machine Learning:​​​‌ error control and analysis,​ Besancon, January 15-16 2025​‌

Claire Monteleoni: Keynote, 25th​​ Anniversary of the Bjerknes​​​‌ Centre for Climate Research,​ Bergen, Norway, March 2025​‌

Claire Monteleoni: College de​​ France, Grand Evênement :​​​‌ L'IA et les mathématiques​ pour la météorologie et​‌ la climatologie, Paris, May​​ 2025

Claire Monteleoni: Keynote,​​​‌ Launch of La Maison​ de l'IA, Université de​‌ Versailles-Saint-Quentin-en-Yvelines, June 2025

Claire​​ Monteleoni: Workshop on AI​​​‌ for the Carbon Cycle,​ CMCC (Euro-Mediterranean Center on​‌ Climate Change), Como, Italy,​​ June 2025

Cécile Mallet,​​​‌ MétéoFrance / Inria joint​ workshop, Toulouse, October 2025​‌

Emmanuel de Bézenac: Inaugural​​ Conference of PAV-IA, University​​​‌ of Pavia, Italy, October​ 2025

Claire Monteleoni: ECCE​‌ (Expertise Center for Climate​​ Extremes) Seminar, University of​​​‌ Lausanne, October 2025

Claire​ Monteleoni: Workshop on Uncertainty​‌ Quantification for Climate Science,​​ Institut Henri Poincaré, November​​​‌ 2025

Claire Monteleoni: Keynote,​ EurIPS Rethinking AI workshop,​‌ Copenhagen, December 2025

11.1.5​​ Leadership within the scientific​​​‌ community

Charantonis:

• Lead, SCAI​ (Sorbonne Center for Artificial​‌ Intelligence) / IPSL (Institut​​ Pierre Simon Laplace) masters​​​‌ internship fellowship program
• Bureau,​ SAMA (Statistics for Analysis,​‌ Modelling and Assimilation), IPSL​​
• Co-organizer, AI4Climate seminars

Monteleoni:​​​‌

• External Advisory Board, ICCS​ (Institute of Computing for​‌ Climate Science), Cambridge University,​​ 2023-
• U.S. National Science​​​‌ Foundation (NSF) Advisory Committee​ for Environmental Research and​‌ Education, 2021-2025
• Advisory Board,​​ Climate Change AI, 2021-​​​‌
• Global Partnership on AI​ (GPAI) Committee on Climate​‌ Action & Biodiversity Preservation​​ 2021-

11.1.6 Research administration​​​‌

Monteleoni: Scientific Committee, IFREMER,​ 2025-

11.2.1 Founding and leadership​ of degree programs

Barthès​‌ and Chazottes (and Co-Founded​​ by Mallet), Head of​​​‌ Master TRIED (ML &​ Data Processing) Neural Networks​‌ /Statistical Analysis of Real​​ Datasets/Applied Artificial Intelligence Univerité​​​‌ UVSQ-Paris-Saclay/ IPP/ CNAM TRIED​

11.2.2 Teaching

Julie Keisler,​‌ Practical Work: Introduction to​​ Deep Learning, Faculté des​​​‌ Sciences d'Orsay - Université​ Paris Saclay

11.2.3 Supervision​‌

HDR Sarah Safieddine. February​​ 2025 FR : Interactions​​​‌ entre la Température et​ la Composition Atmosphérique de​‌ la Surface à la​​ Stratosphère ENG: Interactions Between​​​‌ Temperature and Atmospheric Composition​ from the Surface to​‌ the Stratosphere

PhD students​​ All PhD students listed​​​‌ in the first section​ are supervised or co-supervised​‌ by ARCHES team members.​​

Thesis defenses

• Julie KEISLER​​​‌ Co-supervisor: Claire Monteleoni Université​ de Lille Discipline: Informatique​‌ Janvier 2025 Optimisation de​​ réseaux de neurones :​​​‌ algorithmes et logiciel pour​ un système électrique durable​‌ Automated Deep Learning :​​ algorithms and software for​​​‌ energy sustainability
• Ganglin TIAN​ Co-supervisor: Anastase Charantonis SORBONNE​‌ UNIVERSITÉ Discipline : Sciences​​ de l'Atmosphère Novembre 2025​​​‌ PRÉVISIONS MÉTÉOROLOGIQUES POUR L'ÉNERGIE​ AUX ÉCHÉANCES SOUS- SAISONNIÈRES​‌ Improving Sub-seasonal Weather Forecasts​​ for Energy

New PhD​​​‌ students in 2025

• Aymeric​ Delefosse, Inria supervisor: Anastase​‌ Charantonis
• Nathan Chalumeau, CIFRE​​ AXA Research, Inria supervisor:​​​‌ Emmanuel de Bezenac
• Renu​ Singh, Google DeepMind, Inria​‌ supervisor: Claire Monteleoni

CSI:​​ Committe de Suivi Individuel​​​‌ Monteleoni:

• Amaury Lancelin, ENS​
• Pierre Chapel, ENS

11.2.4​‌ Juries

• Mallet:
  • Rapporteure, Thesis​​ defended by Raul Carreira​​ Rufato "Reconnaissance des arcs​​​‌ de défaut dans les‌ aéronefs par apprentissage artificiel"‌​‌ - École doctorale :​​ Informatique, Télécommunications et Électronique​​​‌ de Paris - Sorbonne‌ université
  • Rapporteure, Thesis defended‌​‌ by Daria Botvynko "Lagrangian​​ trajectories simulation on the​​​‌ sea surface using deep‌ Learning" - École doctorale‌​‌ : Mathématiques et Sciences​​ et technologies de l'Information​​​‌ et de la Communication‌ en Bretagne Océane
  • Rapporteure,‌​‌ Thesis defended by Clément​​ Bazantay "Quantification of ice​​​‌ crystal morphological properties in‌ deep convective cloud systems,‌​‌ based on in-flight observations"​​ Ecole doctorale des Sciences​​​‌ Fondamentales
• Monteleoni:
  • Tenure committee‌ member, Tom Buecler, University‌​‌ of Lausanne, October 2025​​
  • Rapporteure, HDR, Dennis Wilson,​​​‌ Université Toulouse Capitole, April‌ 2025
  • Examiner, HDR, Pierre‌​‌ Gaillard, Université Grenoble Alpes,​​ July 2025
  • Chair, Thesis​​​‌ defense, Lawrence Stewart, ENS,‌ June 2025
  • Examiner, Ganglin‌​‌ Tian (see Thesis defenses​​ above), November 2025

11.3.1 Productions (articles,‌ videos, podcasts, serious games,‌​‌ ...)

Safieddine: Article in​​ the Conversation France: link​​​‌

11.3.2 Participation in Live‌ events

Monteleoni: Invited panelist,‌​‌ Global Talent, French Future:​​ Stories of AI Researchers​​​‌ in France. Vivatech, Paris,‌ June 2025

Monteleoni: Invited‌​‌ panelist, L'apport de l'IA​​ dans l'adaptation au changement​​​‌ climatique, round table organized‌ by CEREMA and the‌​‌ Société d'Encouragement pour l'Industrie​​ Nationale, Paris, July 2025​​​‌

11.3.3 Others science outreach‌ relevant activities

ARCHES research‌​‌ mentioned in the media:​​

• Le Monde Science &​​​‌ Medecine, Pour une convention‌ nationale de I'éthique environnementale‌​‌ de la recherche. 29​​ Jan 2025 P.7
• L'Info​​​‌ Durable, L'IA peut-elle vraiment‌ aider à lutter contre‌​‌ le réchauffement climatique ?​​. 12 Feb 2025​​​‌
• CHOISEUL MAGAZINE L'IA, le‌ levier qui accélérera la‌​‌ transition environnementale Edition :​​ Winter 2025 P.24-27

International​​​‌ journals

• 19 articleC.‌Clément Dauvilliers and C.‌​‌Claire Monteleoni. MoTiF:​​ a self-supervised model for​​​‌ multi-source forecasting with application‌ to tropical cyclones.‌​‌Environmental Data Science4​​July 2025, e36​​​‌HALDOI
• 20 article‌D.David Landry,‌​‌ C.Claire Monteleoni and​​ A. A.Anastase Alexandre​​​‌ Charantonis. Generating ensembles‌ of spatially coherent in‌​‌ situ forecasts using flow​​ matching.Quarterly Journal​​​‌ of the Royal Meteorological‌ SocietyOctober 2025HAL‌​‌DOI
• 21 articleG.​​Ganglin Tian, A.​​​‌ A.Anastase Alexandre Charantonis‌, C. L.Camille‌​‌ Le Coz, A.​​Alexis Tantet and R.​​​‌Riwal Plougonven. Quantile‌ Regression, Variational Autoencoders, and‌​‌ Diffusion Models for Uncertainty​​ Quantification: A Spatial Analysis​​​‌ of Sub-seasonal Wind Speed‌ Prediction.Monthly Weather‌​‌ Review2025, 415​​ - 422HAL
• 22​​​‌ articleG.Ganglin Tian‌, C. L.Camille‌​‌ Le Coz, A.​​Alexis Tantet, A.​​​‌Anastase Charantonis, N.‌Naveen Goutham and R.‌​‌Riwal Plougonven. Improving​​ Subseasonal Wind Speed Forecasts​​​‌ in Europe with a‌ Nonlinear Model.Monthly‌​‌ Weather Review1539​​September 2025, 1761-1779​​​‌HALDOI
• 23 article‌M.Maurin Zouzoua,‌​‌ S.Sophie Bastin,​​ F.Fabienne Lohou,​​​‌ M.Marie Lothon,‌ M.Marjolaine Chiriaco,‌​‌ M.Mathilde Jome,​​ C.Cécile Mallet,​​​‌ L.Laurent Barthes and‌ G.Guylaine Canut.‌​‌ Using a data-driven statistical​​ model to better evaluate​​​‌ surface turbulent heat fluxes‌ in weather and climate‌​‌ numerical models: a demonstration​​ study.Geoscientific Model​​​‌ Development1811June‌ 2025, 3211-3239HAL‌​‌DOI

Conferences without proceedings​​

• 24 inproceedingsM.Matthieu​​​‌ Meignin, N.Nicolas‌ Viltard, L.Laurent‌​‌ Barthes and C.Cécile​​ Mallet. Refining Infrared-Only​​​‌ Rainfall Estimation with Deep‌ Learning.Climate Informatics‌​‌ 2025Rio de Janeiro,​​​‌ BrazilApril 2025HAL​
• 25 inproceedingsN.Nicolas​‌ Viltard, V.Vibolroth​​ Sambath, A.Audrey​​​‌ Martini, L.Laurent​ Barthès and C.Cécile​‌ Mallet. Evolution of​​ global rain intensities over​​​‌ the TRMM and GPM​ era using a deep-learning​‌ algorithm to assess the​​ impact of climate change​​​‌ on Earth water cycle​.Climate Informatics 2025​‌Rio de Janeiro, Brazil​​April 2025HAL

Edition​​​‌ (books, proceedings, special issue​ of a journal)

• 26​‌ proceedingsAutoML algorithms for​​ online generalized additive model​​​‌ selection: application to electricity​ demand forecasting.AutoML​‌ 2025: International Conference on​​ Automated Machine Learning293​​​‌New York (NY), United​ StatesPMLRNovember 2025​‌, 23/1--19HAL

Reports​​ & preprints

• 27 misc​​​‌G.Graham Clyne,​ G.Guillaume Couairon,​‌ G.Guillaume Gastineau,​​ C.Claire Monteleoni and​​​‌ A.Anastase Charantonis.​ ArchesClimate: Probabilistic Decadal Ensemble​‌ Generation With Flow Matching​​.2026HALback​​​‌ to text
• 28 misc​J.Julie Keisler,​‌ A. A.Anastase Alexandre​​ Charantonis, Y.Yannig​​​‌ Goude, B.Boutheina​ Oueslati and C.Claire​‌ Monteleoni. SerpentFlow: Generative​​ Unpaired Domain Alignment via​​​‌ Shared-Structure Decomposition.December​ 2025HALback to​‌ text