Thèse, Paris.| Entreprise/Organisme : | Sorbonne Universite, Institut des Systèmes Intelligents et de Robotique (ISIR) | | Niveau d'études : | Master | | Sujet : | Abstract: AI4Science is an emerging field investigating the potential of AI to advance scientific discovery, with deep learning playing a central role in modeling complex natural phenomena. Within this context, deep generative modeling—which already enables the synthesis of high-dimensional data across modalities such as text, images, and audio—is now opening new avenues for simulating and understanding complex physical systems.
This PhD project aims to explore and advance generative deep learning architectures—such as diffusion models, flow-matching networks, and autoregressive transformers—for modeling complex physical dynamical systems arising in domains such as climate, biology, and fluid mechanics. These models hold strong potential for learning flexible, data-driven representations of physical laws. By developing generalizable, cross-physics generative models, this research contributes to the broader vision of AI4Science: accelerating scientific discovery through learned simulation and abstraction. | | Date de début : | November or December 2025 | | Durée du contrat : | 36 mois | | Rémunération : | 2200 per Month Gross Salary + possible teaching vacations | | Secteur d'activité : | Computer Science, Artificial Intelligence, Machine Learning | | Description : | This PhD project aims to explore and advance generative deep learning architectures—such as diffusion models, flow-matching networks, and autoregressive transformers—for modeling complex physical dynamical systems arising in domains such as climate, biology, and fluid mechanics. These models hold strong potential for learning flexible, data-driven representations of physical laws. By developing generalizable, cross-physics generative models, this research contributes to the broader vision of AI4Science: accelerating scientific discovery through learned simulation and abstraction.
Research Objectives
The overarching research question is: Can we develop generative models that learn structured, physically grounded representations of dynamical systems—enabling synthesis, adaptation, and generalization across physical regimes and multiphysics settings? It unfolds into several complementary directions:
Latent Generative Models for Physical Dynamics
The objective is to design generative models—such as diffusion, flow-matching, or autoregressive models—that learn compact and interpretable latent representations of spatiotemporal dynamics governed by PDEs. These models should:
• Capture uncertainty and multimodality in solution trajectories.
• Generalize across parametric variations.
Learning Across Multiphysics Systems
To enable transfer learning across heterogeneous physics, we will explore shared latent representations across families of PDEs:
• Using encode–process–decode frameworks.
• Applying contrastive or multitask training to uncover reusable physical abstractions.
• Designing models invariant to space/time resolution and units.
This direction builds toward foundation-like models that capture generalizable physics priors across simulation families.
Few-Shot and In-Context Generalization to New Physics
To support scientific modeling in data-scarce settings, we will develop methods for few-shot generalization such as:
• Fine-tuning latent priors to new PDE systems using limited examples.
• Exploring meta-learning and prompt-based adaptation techniques (inspired by in-context learning in language models).
• Incorporating known physical constraints into the generative process.
The goal is to enable rapid and physically consistent adaptation to previously unseen dynamics with minimal data and supervision.
Position and Working Environment
The PhD studentship is a three years position starting in October/November 2025. It does not include teaching obligation, but it is possible to engage if desired. The PhD candidate will work at Sorbonne Université (S.U.), in the center of Paris. He/She will integrate the MLIA team (Machine Learning and Deep Learning for Information Access) at ISIR (Institut des Systèmes Intelligents et de Robotique).
References
Chen, W., Song, J., Ren, P., Subramanian, S., Morozov, D., & Mahoney, M. W. (2024). Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning. 1–21. http://arxiv.org/abs/2402.15734
Hao, Z., Su, C., Liu, S., Berner, J., Ying, C., Su, H., Anandkumar, A., Song, J., & Zhu, J. (2024). DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Training. Icml. http://arxiv.org/abs/2403.03542
Kassai Koupai, A., Benet, J. M., Yin, Y., Vittaut, J.-N., & Gallinari, P. (2024). GEPS: Boosting Generalization in Parametric PDE Neural Solvers through Adaptive Conditioning. NeurIPS. https://geps-project.github.io/
Kirchmeyer, M., Yin, Y., Donà, J., Baskiotis, N., Rakotomamonjy, A., & Gallinari, P. (2022). Generalizing to New Physical Systems via Context-Informed Dynamics Model. ICML.
McCabe, M., Blancard, B. R.-S., Parker, L. H., Ohana, R., Cranmer, M., Bietti, A., Eickenberg, M., Golkar, S., Krawezik, G., Lanusse, F., Pettee, M., Tesileanu, T., Cho, K., & Ho, S. (2024). Multiple Physics Pretraining for Physical Surrogate Models. 1–25 http://arxiv.org/abs/2310.02994
Serrano, L., Wang, T., le Naour, E., Vittaut, J.-N., & Gallinari, P. (2024). AROMA : Preserving Spatial Structure for Latent PDE Modeling with Local Neural Fields. NeurIPS.
Serrano, L., Kassai, A., Wang, T., Erbacher P., Gallinari, P., (2025) Zebra: In-Context Generative Pretraining for Solving Parametric PDEs.
Zhou, A., Li, Z., Schneier, M., Buchanan Jr, J. R., & Farimani, A. B. (2025). TEXT2PDE: Latent Diffusion Models for Accessible Physics Simulation. ICLR. | | En savoir plus : | https://pages.isir.upmc.fr/gallinari/open-positions/
2025-05-01-PhD-Description-Generative-models-Physics.pdf | | Contact : | patrick.gallinari@sorbonne-universite.fr |
|
