Tadpole: TUM uses autoencoders as foundation models for 3D PDEs with online learning
jiqizhixin · x · 2026-10-23
Technical University of Munich presents Tadpole (NeurIPS, ArXiv 2605.15284), tackling a key obstacle for physics foundation models: every 3D physical system has its own variables, resolution and evolution laws, and high-res fluid simulation data is hugely expensive to store.
Instead of learning to "predict the next step" first, Tadpole pretrains autoencoders as foundation models for 3D PDEs with online learning: it trains on single-channel crops generated on-the-fly by a GPU solver, using an efficient buffer strategy to eliminate I/O and storage bottlenecks. The pretrained model supports multiple downstream tasks — autoencoding, dynamics learning via the novel Tadpole-DFT method (LoRA fine-tuning, latent transformation, skip connections), and generative modeling via latent flow matching — reframing scientific foundation models from next-step predictors into world-understanders.
More from Research
- AnyGoal research asks whether world models memorize trajectories instead of learning reusable rules — randall_balestr · 2026-10-23
- Stratified Sparse Attention unifies retrieval and sampling to close the sparse-to-dense gap at extreme sparsity — Skylight1 · 2026-10-23
- New paper with 25 coauthors examines feasibility of a hardwired pause on frontier AI training — wfithian · 2026-10-23
- FDA crosses a regulatory watershed with personalized neoantigen cancer vaccine, per Gottlieb — alex_peys · 2026-10-23
- VIPerturb-seq Update: 10x Genomics to Commercialize Single-Cell CRISPR Screening Tech — anshulkundaje · 2026-10-23
- New Paper Replicates Transformer Loss Spikes with Simple Quadratic Models — CSProfKGD · 2026-10-23