Internship: Active Learning for Fundational PDE Surrogates

Inria Saint Martin d'Hères Saint Martin d'Hères, France

About the position

Context

This master internship (M2 or engineering school PFE) will take place at the DataMove team located in the IMAG building on the campus of Saint Martin d’Heres (Univ. Grenoble Alpes) near Grenoble. The length of the internship is
4 months minimum and the start date is flexible, but need a 2 month delay before starting the interhsip due to administrative constraints. The city of Grenoble is a student-friendly city surrounded by the Alps mountains, offering a high quality of life and where you can experience all kinds of mountain-related outdoor activities.
Advisers: Bruno Raffin (
Bruno.Raffin@inria.fr
)

Assignment

Context

A
PDE surrogate is a neural network trained to approximate a classical PDE solver — for instance, a CFD solver. The solver acts as a data generator that produces the training set.
An extensive line of work has focused on designing neural architectures and training procedures that yield the best possible neural surrogate. One trend is to make surrogates as generic as possible — for example, able to handle different initial and boundary conditions, different meshes, and so on.
Foundational PDE surrogates push this logic even further, training a single surrogate on data drawn from many different PDEs. The standard neural architectures used for these models are often derived from vision networks such as vision transformers.
Our approach: online training and OGAS
Our team has been developing online training for PDE surrogates, in which PDE solver instances are executed to generate data that is used directly to train the surrogate (
Meyer et al., ICML 2023
). The first benefit is faster training (a 10× speed-up is common on supercomputers), but online training also makes it possible to steer data generation dynamically based on the progress of training, thereby improving the generalization of the trained surrogate.
We developed such an active learning method, called
OGAS
(
Cesar at al., Neurips 2026
). OGAS trains a fast diffusion model in parallel with the surrogate to learn the probability distribution of samples solver parameters conditioned on a training-difficulty signal such as the loss. Each time a new solver instance must be run, the conditional sampler is used to draw the solver's input parameters, so as to generate hard-to-learn training data. Experiments demonstrated significant gains across different PDEs and models.

Main activities

Internship objectives
The goal of this internship is to extend this approach along two axes:
F oundational PDE surrogates
.
Instead of training a surrogate for a single PDE, the goal is to train one surrogate for several PDEs, with the expectation that it can be quickly fine-tuned to learn an unseen PDE. Extending OGAS to this setting will require tackling the issue of data heterogeneity
​.
Trajectory stability
.
A common issue for all PDE surrogates is the stability of the generated trajectories: the recurrent application of the surrogate to its own previous output tends to accumulate errors, leading to degraded output quality over long roll-outs. Recent papers have investigated methods to gain stability, mainly by injecting noise into the training data. We will investigate how to integrate and extend these approaches within OGAS.
Environment
We maintain a full software suite for online surrogate training, called
Melissa
​
, which includes OGAS, various PDE solvers, and neural models. The intern will leverage this existing software base to focus quickly on developing her/his own contribution. The intern will also have access to GPU supercomputers for experiments.
Related material
OGAS:
https://melissa-sa.github.io/ogas/
Melissa:
https://melissa-sa.github.io/
Foundational PDE surrogate papers:
Poseidon — http://arxiv.org/abs/2405.19101
PDE-Transformer — https://arxiv.org/abs/2505.24717
Transolver — http://arxiv.org/abs/2602.04940
AB-UPT — http://arxiv.org/abs/2502.09692
Trajectory stability paper:
http://arxiv.org/abs/2609.32864

Skills

The main communication language is English.

This listing was collected from a public source and is reproduced here for information only. Always confirm the details on the original posting before applying.
View the original posting

Similar positions