PhD Position F/M PhD Position F/M - Scalable and geometry-aware AI architectures for weather forecasting from satellite observations

Contexte et atouts du poste

This PhD thesis will be directed by Claire Monteleoni and co-supervised by Emmanuel de Bézenac within the ARCHES project-team at Inria Paris. Emmanuel de Bézenac will be the primary day-to-day scientific supervisor. Dino Ienco and Diego Marcos, from the EVERGREEN project-team, will also participate in the broader supervisory team, providing complementary scientific guidance and regular exchanges throughout the PhD.

The position is part of the INRIA-CNES GRASP challenge, a joint research initiative dedicated to developing scalable and transferable foundation models for large, heterogeneous and sparsely labelled Earth-observation archives. The PhD student will also collaborate closely with researchers and Earth-observation experts from CNES.

This supervisory environment brings together complementary expertise in machine learning for weather and complex dynamical systems, multimodal remote sensing, geometric and equivariant deep learning, satellite observations and high-performance computing. The student will benefit from regular interactions across the participating Inria teams and CNES, while being primarily based within the ARCHES team at Inria Paris.

Recent advances in artificial intelligence have demonstrated that data-driven models can achieve highly competitive performance in short- and medium-range weather forecasting. However, most existing approaches rely on a relatively small number of atmospheric variables represented on fixed and regularly sampled grids. They are therefore not designed to fully exploit the rapidly growing diversity of satellite and Earth-observation data.

Satellite instruments provide measurements with different physical meanings, spatial resolutions, temporal frequencies, sampling geometries and uncertainty levels. Observations may also be irregular, partially missing or available only over relatively short historical periods. Designing models that can jointly process these heterogeneous sources, scale to very large numbers of variables and observations, and remain effective when training data are limited constitutes a major scientific challenge.

Mission confiée

The objective of this PhD is to develop generic, scalable and geometry-aware machine-learning architectures for weather forecasting from heterogeneous satellite observations.

The research will investigate how a common model can ingest and combine very large numbers of variables that differ in their physical nature, coordinate systems, spatial and temporal resolutions, sampling geometries and availability. Rather than assuming that every variable is represented on the same predefined grid, the PhD will explore flexible architectures capable of processing observations across different sensors, locations and scales.

A central challenge will be to obtain architectures that are sufficiently generic to accommodate new variables and observation systems while maintaining strong forecasting performance. The models should be capable of scaling to large and diverse inputs without requiring an unreasonable growth in computational cost or model size.

The thesis will also investigate how to learn efficiently when the amount of available training data is limited. This is particularly important for recently launched satellite instruments, variables with short historical records, poorly observed geographical regions and rare atmospheric phenomena. Relevant directions may include self-supervised learning, multimodal pretraining, parameter sharing, transfer learning and the introduction of appropriate architectural priors.

Another major research direction will concern the geometry of Earth-observation and weather data. The candidate will study how relative spatial relationships, spherical geometry, changes of resolution and relevant symmetry transformations can be encoded in neural architectures. Equivariance to rotations, translations, changes of orientation or scale may enable models to generalize more effectively across geographical regions, sensors and resolutions.

More broadly, the thesis may investigate how geometric, structural and physical knowledge can guide learning without unnecessarily restricting the expressivity of the model. The overall goal is to design architectures that combine flexibility, scalability, robustness and data efficiency.

Principales activités

The PhD student will conduct research on several of the following topics:

• Design flexible neural architectures for weather forecasting from heterogeneous atmospheric and satellite observations.
• Develop representations that can accommodate very large numbers of variables with different physical meanings, resolutions and sampling structures.
• Investigate alternatives to rigid grid-based tokenization and patchification, including attention-based architectures and models with adaptive computation.
• Develop mechanisms for combining data observed at different spatial and temporal scales.
• Address irregular sampling, missing observations, varying sensor availability and differences between satellite acquisition geometries.
• Study architectures whose inputs and outputs are not tied to a particular grid, patch size, sensor or resolution.
• Investigate self-supervised, multimodal and transfer-learning approaches for improving performance in limited-data regimes.
• Introduce geometric priors and equivariance properties relevant to weather forecasting and Earth observation.
• Explore relative spatial representations, spherical geometries and multiscale modelling.
• Investigate soft physical or structural constraints that could improve the consistency and robustness of forecasts.
• Develop computational strategies for scaling the proposed models to extremely large numbers of observations and variables.
• Compare the resulting approaches with existing data-driven weather forecasting architectures.
• Evaluate predictive accuracy, robustness, transfer across sensors and geographical regions, data efficiency and computational cost.
• Conduct ablation and scaling studies to understand the respective contributions of the proposed architectural components.
• Present research results at international conferences and publish them in leading machine-learning, remote-sensing or weather and climate venues.
• Contribute to the development of reproducible research software and, when possible, open-source implementations.

The precise research directions will be refined during the PhD according to the candidate's interests, the available datasets and the initial experimental results.

Compétences

Candidates should hold, or be close to obtaining, a Master's degree or equivalent in machine learning, computer science, applied mathematics, statistics, physics, atmospheric science or a related discipline.

The following skills are expected:

• Strong foundations in machine learning and deep learning.
• A good mathematical background, particularly in linear algebra, probability, optimization or numerical methods.
• Experience with a modern deep-learning framework such as PyTorch or JAX.
• Good programming skills and familiarity with scientific computing workflows.
• Ability to design and evaluate machine-learning experiments rigorously.
• Ability to work with large and potentially complex scientific datasets.
• Strong interest in at least some of the following areas: weather forecasting, foundation models, geometric deep learning, equivariant neural networks, spatiotemporal modelling, remote sensing or multimodal learning.

Experience with transformers, distributed training, high-performance computing, geospatial data, satellite observations or atmospheric datasets would be appreciated but is not required.