Internship In Context Learning For Surrogate Models In Physics Simulations H/F Safran

Context
Deep learning surrogates such as GNNs, mesh-based transformers, and reduced-order models are increasingly used to accelerate physics simulations, but require significant task-specific training and engineering. Recently, foundation models for tabular data (e.g. TabPFN) have been proposed as training-free models that adapt to new tasks via conditioning only. Their relevance for physics simulation data remains largely unexplored.
SafranTech has already benchmarked multiple surrogate modeling approaches within the PLAID framework (https://huggingface.co/PLAIDcompetitions), providing a solid reference for comparison.

Objectives
The internship aims to:
- Assess whether foundation models for tabular data can serve as viable surrogates for physics simulations using conditioning only.
- Compare their accuracy and robustness with existing deep learning methods benchmarked in PLAID.
- Evaluate their computational throughput and scalability for large-scale industrial use.

Work Plan
1. Bibliography and Benchmark Setup
Review foundation models for tabular data and physics-based surrogate modeling. Become familiar with PLAID datasets, metrics, and baselines.
2. Regular Grid Experiments
Apply foundation models directly to physics simulation outputs defined on regular grids. Compare accuracy and inference cost with existing PLAID baselines.
3. Throughput and Scalability Analysis
Benchmark inference time, memory usage, and batch scalability, and compare conditioning-only inference with fully trained surrogates.
4. Unstructured Grids via Latent Representations
Extend the approach to unstructured meshes using latent embeddings (e.g. POD or autoencoders), and compare with mesh-aware models.
5. Analysis and Limitations
Identify failure modes, scalability limits, and applicability domains for foundation models in physics simulations.

Expected Outcomes
- A quantitative comparison between foundation models and state-of-the-art physics surrogates.
- Clear conclusions on accuracy-throughput trade-offs.
- Practical recommendations for industrial simulation workflows.