Internship: Online Simulation-based Inference for Large-scale Scientific Models

Le descriptif de l’offre ci-dessous est en Anglais_

**Type de contrat**: Convention de stage

**Niveau de diplôme exigé**: Bac + 4 ou équivalent

**Fonction**: Stagiaire de la recherche

**A propos du centre ou de la direction fonctionnelle**:
The Centre Inria de l’Université de Grenoble groups together almost 600 people in 23 research teams and 9 research support departments.

Staff is present on three campuses in Grenoble, in close collaboration with other research and higher education institutions (Université Grenoble Alpes, CNRS, CEA, INRAE,), but also with key economic players in the area.

The Centre Inria de l’Université Grenoble Alpe is active in the fields of high-performance computing, verification and embedded systems, modeling of the environment at multiple levels, and data science and artificial intelligence. The center is a top-level scientific institute with an extensive network of international collaborations in Europe and the rest of the world.

**Contexte et atouts du poste**:
The length of the internship is _4 months minimum_ and the start date is flexible, but need a 2 months delay before starting the interhsip due to administrative constraints. The DataMove and STATIFY teams are friendly and stimulating environment that gathers Professors, Researchers, PhD and Master students all leading research on High-Performance Computing and Machine Learning. 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.

**Mission confiée**:
**Context**

Researchers are turning to machine learning to tackle various problems in science, from biology to astro-physics and fluid dynamics. The project that we propose is part of this growing AI4Science movement, focusing on a key challenge in experiments: figuring out which model parameters best match the data we observe (Figure 1). More specifically, we use simulation-based inference (SBI) [1], a Bayesian approach that leverages deep generative models, such as conditional normalizing flows and score-diffusion models, to approximate the posterior distribution assigning higher probability to parameter values most likely to have produced an observed data.

**Principales activités**:
When considering large scale simulations the amount of data produced can be overwhelming and the execution time too long, calling to resort to supercomputers and High Performance Computing (HPC). To optimize performance and reduce costs (power, storage, compute time), training can be performed online. Multiple simulations are executed concurrently and continuously to produce data that are used asap, without being stored in files, by a training process that also runs concurrently with these simulations (Fig. 2) [2, 3]. The traditional SBI workflow consisting of (simulate store, then store train) has to be re-visited to properly support and leverage this online training workflow (simulate buffer train).

To be more specific, consider the usual SBI approach of training a conditional neural density estimator qϕ that approximates the target posterior through the minimization of

The batch of N pairs of parameters (θi) and simulations (xi) is provided upfront and the loss function is minimized via some variant of stochastic gradient descent. Note that this is well motivated because when N, one can show that the minimizer of Equation 1 is indeed the target posterior p(θ|x). However, it is not clear how the minimization behaves when the training samples are obtained sequentially due to simulation constraints and/or sampled from a different distribution than p(θ,x) as one would do when trying to reduce the number of calls to the simulator.
- What is the direct impact of an online paradigm for simulations on the usual batched SBI training? What are the precise bottlenecks and challenges to this transition?
- Are there other loss functions more appropriate to minimize instead of Equation 1 when working under the paradigm of large-scale simulators?
- Can approaches from online reinforcement learning to make smart queries to the simulator and minimize total cost of compute can be reused and adapted?

**References**

[1] MichaelDeistler, JanBoelts, PeterSteinbach, GuyMoss, ThomasMoreau, ManuelGloeckler, PedroLCRodrigues, Julia Linhart, Janne K Lappalainen, Benjamin Kurt Miller, et al. Simulation-based inference: A practical guide. arXiv preprint arXiv:2508.12939, 2025.

[2] Sofya Dymchenko and Bruno Raffin. Loss-driven sampling within hard-to-learn areas for simulation-based neural network training. In MLPS 2023-Machine Learning and the Physical Sciences Workshop at NeurIPS 2023-37th conference on Neural Information Processing Systems, pages 1-5, 2023.

[3] LucasThibautMeyer,MarcSchouler,RobertAlexanderCaulk,AlejandroRibés,andBrunoRaffin. Highthroughput training of deep surrogates from large ensemble runs. In Proceedings of the International Conference for High