Research
My research combines data assimilation, scientific machine learning and reduced-order modelling to reconstruct, predict and control complex physical systems from limited observations.
My main interests include latent-space data assimilation, physics-aware representation learning, generative models and differentiable simulation, with applications to wildfire, atmospheric science and fluid dynamics.
Research Projects
Hi! PARIS Starting Fellowship
Differentiable Digital Twin for Wildfire Prediction & Control
Role: Principal Investigator
Programme: Hi! PARIS Internal Chairs / Fellowships, 2026 cohort
This project develops a differentiable digital twin for wildfire prediction and control, connecting physical modelling with machine learning. The research focuses on differentiable wildfire simulation, calibration and prediction to support effective fire management.
OFML — Machine Learning for Wildfire Fighting Strategies
Optimizing Real-Time Wildfire Fighting Strategies Using Machine Learning Techniques
Role: Principal Investigator
Funding: ANR — ANR-25-CE56-0198
Period: 2025–2029
OFML develops efficient wildfire prediction models that account for human interventions, including firebreaks and retardant deployment. It combines GPU-based simulation, deep learning and reinforcement learning to optimise firefighting decisions, with a focus on France and the Mediterranean region. I coordinate the project and lead its research on efficient wildfire modelling and optimisation of firefighting strategies.
MilaDy — Machine Learning for Turbulent Dynamos
Role: Co-Investigator
Project Coordinator: Caroline Nore, LISN, Université Paris-Saclay
Funding: ANR — ANR-25-CE51-0077
Start: 2026
MilaDy combines machine learning and numerical simulation to study turbulent magnetohydrodynamic dynamos: conducting fluid flows that generate magnetic fields. The project investigates magnetic-field reversals, rare dynamical events and the influence of physical parameters.
At CEREA/ENPC, our research focuses on physics-aware representation learning and reduced-order modelling of three-dimensional velocity and magnetic fields, connecting learned representations to data assimilation.
QWERTY — Forest Ecological Vulnerability to Wildfires
Quantifying the Vulnerability of Forest Ecological Values to Wildfires
Role: Co-Investigator
Project Coordinator: Florent Mouillot, IRD
Funding: France 2030 — PEPR FORESTT, ANR-25-PEFO-0008
Period: 2025–2029
QWERTY assesses wildfire impacts beyond burned area and biomass loss. It combines ecological observations, satellite data and modelling to map forest ecological values, evaluate vulnerability and post-fire recovery, and support ecological risk assessment.
At CEREA/ENPC, we develop GPU-accelerated wildfire modelling with PyTorchFire, supporting the assessment of combustion and greenhouse-gas emissions within WP3.
Related Publications
See the Publications page for research papers related to these projects.