Etienne Meunier

I am a Postdoc at Inria Ange Team in Paris, where I work closely with LOCEAN-IPSL on machine learning methods for physical systems, with a focus on ocean and climate modelling. I came to climate science after a PhD on unsupervised training of neural networks for motion analysis, which I did in the Serpico team supervised by Patrick Bouthemy. Before that, from 2019 to 2020, I was a visiting scholar in the Mostafavi Lab at the University of British Columbia, where I developed deep learning approaches to interpret genomics data.

My interests lie where machine learning meets physics. I like building models that learn from data while still respecting the equations behind it, whether they are hybrid models that stay stable over long simulations, generative models that produce physically consistent ocean states, or differentiable simulators. Along the way, I enjoy working with climate scientists and writing research code that others can actually use.

At Inria / IPSL, I co-supervise the PhDs of Blandine Gorce, who builds generative models to find stable states of large-scale partial differential equations, and Maya Janvier, who works on making hybrid models stable. I also work closely with Rares Dimitrie Grozavescu at the University of Cambridge on Koopman operators, and with Markus Jochum at the Niels Bohr Institute in Copenhagen on differentiable ocean models.

If you work on related topics, or you are a student looking for an internship, feel free to reach out!

I also have a Blog where I write about the things I am interested in and wish I had more time to work on!

etienne.meunier@inria.fr  /  CV  /  Phd-Manuscript  /  GitHub  /  Google Scholar  /  LinkedIn

profile photo

Research

My current work focuses on machine learning for physical systems and ocean modelling. Before that, I worked on unsupervised video analysis. Today I mainly work on three things:

  • Hybrid models. Coupling neural networks with numerical models, and making sure they stay stable over long simulations.
  • Generative models for the ocean. Producing physically consistent ocean states, for example to speed up the spin-up of ocean models.
  • Differentiable and learned dynamics. Differentiable ocean models and learned latent dynamics such as Koopman operators, for forecasting and data assimilation.
project image Hover Image

Invertible continuous latent dynamic for long-term data assimilation in complex physical systems


Rares Grozavescu*, Etienne Meunier*, Pengyu Zhang and Mark Girolami
Submitted to ICLR 2027, 2026
arxiv /

We propose a continuous-time Koopman autoencoder whose linear latent dynamics can be integrated in closed form at any horizon in a single step. It gives stable, fast long-term forecasts (110x faster than diffusion baselines) and makes data assimilation through the emulator cheap and accurate, with a cost independent of the assimilation window.
project image Hover Image

Unrolled gradients in disguise: bridging interpolation-based and Jacobian regularization for stable neural dynamics


Maya Janvier and Etienne Meunier
NeurIPS 2026, 2026
paper / arxiv /

We show that training neural dynamics on interpolated or noisy samples and regularizing their Jacobian both approximate the gradients of unrolled training. Building on this, we derive a principled regularization framework that stabilizes long-term integration of learned dynamical systems without backpropagating through the solver.
project image Hover Image

Divergence-Free Neural Networks with Application to Image Denoising


Sébastien Herbreteau and Etienne Meunier
ICLR 2026, 2026
paper / code /

We introduce a resource-efficient neural network architecture with zero divergence by design, adapted to high-dimensional problems. Applied to self-supervised image denoising through Stein’s unbiased risk estimation, it is competitive with other divergence-based approaches and outperforms them when the noise level is unknown and varies across the training data.
project image Hover Image

Physically Consistent Sampling For Ocean Model Initialization


Blandine Gorce, Luther Ollier, David Kamm and Etienne Meunier
NeurIPS Workshop : Tackling Climate Change with Machine Learning, 2025
paper / code /

We use a diffusion-based ocean emulator to accelerate the costly spin-up phase of ocean models, and introduce physically meaningful constraints on vertical stratification to guide sampling toward physically coherent states.
project image Hover Image

Towards fully differentiable neural ocean model with Veros


Etienne Meunier, Said Ouala, Hugo Frezat, Julien Le Sommer and Ronan Fablet
EurIPS Workshop : Differentiable Systems and Scientific Machine Learning, 2025
paper / arxiv / code /

We make the VEROS ocean model fully differentiable with JAX, enabling automatic differentiation through its dynamical core. We check the numerical consistency of the gradients and demonstrate two applications: correcting an initial ocean state and calibrating unknown physical parameters directly from model observations.
project image Hover Image

Learning to generate physical ocean states: Towards hybrid climate modeling.


Etienne Meunier, David Kamm, Guillaume Gachon, Redouane Lguensat and Julie Deshayes
Accepted at ICLR Workshop : Tackling Climate Change with Machine Learning , 2025
arxiv / code /

We learn to generate oceanic states using a diffusion model and to integrate physical constraints during sampling.
project image Hover Image

Unsupervised motion segmentation in one go: Smooth long-term model over a video


Etienne Meunier and Patrick Bouthemy
ArXiv, 2023
arxiv /

We learn to perform motion segmentation over a video sequence in a single pass in a fully unsupervised manner, using a long-term spatio-temporal model based on splines. The method is very fast at test time and provides temporally consistent labels.
project image Hover Image

Unsupervised space-time network for temporally-consistent segmentation of multiple motions


Etienne Meunier and Patrick Bouthemy
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023
paper / video / code /

We extend unsupervised motion segmentation to longer sequences by leveraging spatio-temporal motion models and 3D networks. Provide consistent labels accros the video and improves network robustness to erroneous optical flow.
project image Hover Image

EM-driven unsupervised learning for motion segmentation


Etienne Meunier, Anaïs Badoual and Patrick Bouthemy
IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2023
paper / arxiv / code /

We introduce a loss function theoretically grounded in the expectation maximisation framework to train networks to segment optical flow fields without any ground truth labels. Easy to set-up, robust to errors and computationally cheap, it can also be used as a regularisation term for any video segmentation project.
project image Hover Image

Unsupervised computation of salient motion maps from the interpretation of a frame- based classification network


Etienne Meunier and Patrick Bouthemy
British Machine Vision Conference (BMVC) , 2021
paper / video /

We detect motion saliency in videos using a weakly-supervised approach based on the gradient-based interpretation of a classification network. Applications on handheld camera videos to distinguish parallax from independent motion and on crowd videos for anomaly motion detection.
project image Hover Image

Interpreting deep learning models in genomics using genetic algorithm


Etienne Meunier, Gherman Novakovsky and Sara Mostafavi
Machine Learning in Computational Biology 2019 (MLCB), 2019

We developed a genetic algorithm to generate diverse DNA sequences that maximize the CNN model’s output, which predicts the cell type specificity of accessible DNA regions. Our tool provide a fast way to extract motifs specific to a given cell type and the results show biological relevance.

Papers Under Review

  • Leveraging differentiable programming for ocean model design and calibration
    Said Ouala, Hugo Frezat, Etienne Meunier, Ronan Fablet and Julien Le Sommer
    A differentiable version of the JAX-based Veros ocean model, used to calibrate model parameters with gradients.

  • Versatile Earth-systems COupleR (VerCOR v.0.4.6): Quick and easy coupling and prototyping of Earth system model components in Python and JAX
    Roman Nuterman, Isak Rockström, Vivek Veluvali, Etienne Meunier, Aster Lei Stoustrup and Markus Jochum
    A differentiable JAX coupler for Earth system model components, with built-in Bayesian and gradient-based tuning.

Other Projects

These include coursework, side projects, and research work that has not been published.

project image

Shark detection


Consulting
2019-10-01

We designed a functional shark detection pipeline that combines deep learning techniques for underwater image-based shark detection and an adversarial framework for domain adaptation. The pipeline incorporates a network of real-time underwater cameras, enabling efficient shark detection and immediate communication of alerts to on-ground sites. The pipeline is undergoing testing by the CSR in La Reunion Island.
project image

Micropolis GAN


BU Advanced Machine Learning and Neural Networks
2019-09-01
code /

We trained a GAN to generate city structures within the virtual environment of the game “Micropolis.” Our primary objective was to evaluate the generator’s ability to produce city layouts that adhere to the game’s specific requirements. To gauge the effectiveness of the generator, we compared the cities it generated with those obtained through a genetic algorithm that was guided by the game’s scoring function.

Design and source code from Jon Barron's website.