Resources / Research
Recent papers
Recent field-surrogate architectures from NeurIPS, ICML and ICLR, 2023 to 2026, in order.
- GINO (NeurIPS 2023): Graph operator to a latent grid, FNO there, graph operator back
- GNOT (ICML 2023): Heterogeneous normalised cross-attention, linear complexity
- Transolver (ICML 2024): Physics-attention over learned slices
- UPT (NeurIPS 2024): Fixed-size latent space; perceiver encoder; latent rollout
- LNO (NeurIPS 2024): Physics-cross-attention into and out of a latent space
- AROMA (NeurIPS 2024): Local neural fields in latent space; diffusion for stable rollouts
- Transolver++ (ICML 2025): Adaptive slices, multi-GPU, million-point meshes
- GAOT (NeurIPS 2025): Multiscale attentional graph encoder + transformer processor
- AB-UPT (TMLR 2025): Anchored, branched decoders; up to 150M cells; divergence-free
- Walrus (arXiv 2025): Cross-domain foundation model for continuum dynamics
- CarBench (arXiv 2025): Benchmark of neural surrogates on high-fidelity 3D car aerodynamics
- GIST (arXiv 2026): Gauge-invariant spectral transformer for interactive aerodynamics
- RETO (arXiv 2026): Rotary-enhanced transformer operator for automotive aerodynamics
- Vehicle-family transfer (arXiv 2026): Adapting automotive aero surrogates to new vehicle families
- AeroJEPA (arXiv 2026): Joint-embedding predictive architecture: predict flow latents, decode on demand; semantic latent space
- Read, Write, Relax (arXiv 2026): Why neural PDE surrogates need both global and local processing
- Conformal aero surrogates (arXiv 2026): Multi-granularity conformal prediction for reliable car-aero operators
Related
- Lineages: Four families of field surrogates, from their founding idea to 2026.
- Scaling to millions of points: Attention grows with the square of the points; latent and linear schemes keep it tractable.
- JEPA: Predict the latent of the flow from geometry and condition; decode a field only when you need one.