Activities

The ANR STREAM project is entering a decisive phase where artificial intelligence becomes the engine that unlocks kilometre‑scale air quality modelling. Traditional chemistry‑transport models such as CHIMERE provide robust physics based simulations, but their computational cost limits their ability to resolve the fine‑scale gradients that matter most for exposure, health impact, and urban policy. STREAM aims to bridge this gap by developing a new generation of AI‑based emulators capable of reproducing CHIMERE’s behaviour at a fraction of the cost, while integrating high‑resolution emissions, land‑use proxies, and urban meteorology.

Within this architecture, Work Package 3 is the scientific core of the project: it transforms deterministic simulations into training material for deep learning models, and it builds the neural engines that will downscale pollutant fields from 10–50 km to 1 km and below. This postdoctoral position is embedded precisely at this frontier. It is designed for a researcher who can navigate between atmospheric science and machine learning, who understands both the constraints of physical consistency and the creative potential of generative architectures. The candidate will join a consortium combining LMD’s expertise in air quality modelling, Lab‑STICC’s leadership in geophysical deep learning, and AMPLISIM’s operational deployment capacity, with international partners in Europe and Asia ready to test the demonstrators.

This post‑doc is therefore not only a technical role: it is a scientific pivot that will shape the project’s ability to deliver high‑resolution air quality forecasts, scenario analysis tools, and a scalable AI‑CHIMERE emulator for global use.

Responsabilities

The post‑doctoral researcher will take the lead on designing and implementing the AI downscaling module of STREAM. Their work will begin with the construction of training datasets derived from CHIMERE simulations and enriched with high resolution emissions and land‑use proxies produced in WP1 and WP2. From there, they will explore and compare multiple neural architectures, CNNs, U‑Nets, transformers, and generative models, to identify those best suited for reproducing pollutant concentration fields at kilometer scale resolution.

A central part of the mission will be to ensure that the downscaled outputs remain physically meaningful. This involves embedding constraints such as positivity, mass conservation, and chemical coherence directly into the model design or loss functions. The researcher will also develop species specific strategies for PM2.5, NO2, and O3, accounting for their distinct spatial patterns and sensitivities to meteorology and emissions.

Beyond model development, the post‑doc will implement patch-based training, dynamic sampling, and domain‑wide learning strategies to handle very large spatial domains, including complex regions in South and Southeast Asia. He will collaborate closely with Lab‑STICC on methodological innovation, with LMD on CHIMERE integration, and with AMPLISIM on GPU‑based deployment and operational testing.

The position also includes scientific dissemination: preparing deliverables, contributing to publications, presenting results in consortium meetings, and supporting the creation of an open‑source codebase that will serve both the atmospheric and AI communities.

Working Environment

The Laboratory of Dynamic Meteorology (LMD) is a joint CNRS research unit, hosted at École Polytechnique (Institut Polytechnique de Paris), École Normale Supérieure (PSL University) and Sorbonne University, and is a partner of École des Ponts. Founded in 1968, the LMD studies climate and the environment for Earth and planetary atmospheres. This is an internationally renowned laboratory with approximately 180 staff members, half of whom are permanent employees (researchers, engineers, and administrative personnel). It also includes around forty doctoral students. The laboratory comprises seven scientific teams, support services (administrative team, IT department, and technical department), and two facilities hosted by the Pierre Simon Laplace Institute (IPSL) research federation (the SIRTA observatory and data center), to which the LMD belongs.

The InTro team, within which the postdoctoral researcher will work, studies the physical and chemical properties of the troposphere and its interfaces. This work is part of the modeling and application activities of the CHIMERE model for its various applications in Asia.

Other benefits

• Public transportation in the Paris region (75% flat-rate employer contribution)
• Teleworking is available (subject to manager approval and application)
• Employer subsidy for the on-site restaurant and cafeteria
• Salary includes social security protection (healthcare, unemployment insurance, pension scheme, paid maternity leaves)

Location: On-site at Ecole Polytechnique (Palaiseau, France) with 2 days / week possible teleworking

Salary: the minimum gross salary (before income tax) is 3100 Euros per month, and depends on experience.

Contact:

Essential

– PhD in machine learning, computer science, geosciences, applied mathematics, or related fields

– Strong experience in neural network architecture design for high‑dimensional geophysical data

– Excellent proficiency in PyTorch and GPU‑based training

– Experience with state-of-the-art deep learning architectures (e.g., U‑Nets, transformers, diffusion models, flow matching….)

– Solid programming skills (Python, Git, HPC workflows)

– Ability to work with large datasets and complex workflows

– Experience communicating with diverse audiences (media, conferences, etc.);

– Excellent project management and coordination skills;

Desirable

– Experience with super‑resolution, downscaling, or physics‑informed ML

– Knowledge of atmospheric sciences, air quality or environmental modelling

– Familiarity with CHIMERE or other Chemistry Transport Models