Time to celebrate our 5 accepted NeurIPS’26 papers. Covering a wide range of exciting topics:
- online-pretraining for foundation models, i.e. Tadpole (it’s been online for a long time by now)
- online-pretraining for unstructured neural operator FMs
- pre-training FMs for temporal stability, as well as fundamental, and closely related topics like
- conflict-free training, e.g. for federated learning (improving ConFIG)
- and learning with data from multiple fidelities.
Details on each of them will, of course, follow soon. For now, great work everyone, time to celebrate 🎉 Full paper titles:
Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers; Luis Medrano-Navarro, Giacomo Baldan, Qiang Liu, Benjamin Holzschuh, Jan Hagnberger, Mathias Niepert, Nils Thuerey
Solver-as-Teacher: Solver-Guided On-Policy Post-Training Framework for PDE Foundation Models; Hao Wei, Shangzhe Li, Nils Thuerey
Tadpole: Autoencoders as Foundation Models for 3D PDEs with Online Learning; Qiang Liu, Felix Koehler, Benjamin Holzschuh, Nils Thuerey
CRAFT: Conflict-Resolved Aggregation for Federated Training; Ziqi Wang, Qiang Liu, Nils Thuerey
Compute-optimal data scaling for neural surrogates via multi-fidelity training; Paul Setinek, Felix Koehler, Fabian Paischer, Nils Thuerey, Johannes Brandstetter
