Physics Foundation Models
-
ETH Zurich Poseidon
Hugging Face: https://huggingface.co/camlab-ethz/Poseidon-T, https://huggingface.co/camlab-ethz/Poseidon-B, https://huggingface.co/camlab-ethz/Poseidon-L
GitHub: https://github.com/camlab-ethz/poseidon
Website: https://camlab-ethz.github.io/poseidon/
Technical report: https://arxiv.org/abs/2405.19101Developer: CAMLab, Seminar for Applied Mathematics, ETH Zurich
Released: May 2024
Variants: T, B, L
Parameters: T: 21M / B: 158M / L: 629M
Resolution: 128x128 grid, 4 channels
Architecture: scOT multiscale operator transformer on a SwinV2 backbone, time-conditioned layer norm for continuous-in-time evaluation, 4 hierarchical stages, patch 4, shifted window 16, ConvNeXt residual path; T: embed 48, depths 4/4/4/4 / B: embed 96, depths 8/8/8/8 / L: embed 192, depths 8/8/8/8
License: CC BY-NC 4.0
Modalities: 2D PDE fields in, 2D PDE fields out; density, horizontal velocity, vertical velocity, pressure
Runs on: T/B: Laptop / L: Laptop, 8GB+ memory
Formats: safetensors float32, PyTorch bin
On disk: T: 83.2MB / B: 631.1MB / L: 2.51GB safetensors -
TUM Tadpole
Hugging Face: https://huggingface.co/thuerey-group/Tadpole
GitHub: https://github.com/tum-pbs/Tadpole
Website: https://ge.in.tum.de/2026/05/18/tadpole-flexible-scientific-foundation-models/
Technical report: https://arxiv.org/abs/2605.15284Developer: Thuerey Group, Technical University of Munich
Released: May 2026
Variants: S, B, L; only B weights released
Parameters: S: 8.8M / B: 38.1M / L: 152.1M
Resolution: pre-trained at 64, 128, 256 and 384 cubed, evaluated to 1024 cubed
Architecture: 3D PDE autoencoder pre-trained on single-channel 64x64x64 crops, P3D hybrid backbone with convolutional stages and a transformer bottleneck, adversarial reconstruction loss; latent compression 16 (S) / 8 (B) / 4 (L); Tadpole-DFT adds LoRA, a latent dynamics sub-network, and zero-initialized skip connections for rollout
License: Apache 2.0
Modalities: 3D PDE fields in, 3D PDE fields out
Runs on: Autoencoding: Laptop / Dynamics: Desktop GPU, NVIDIA only
Formats: safetensors, separate encoder and decoder
On disk: B: 60.4MB encoder, 92.6MB decoder -
Polymathic AI Walrus
Hugging Face: https://huggingface.co/polymathic-ai/walrus
GitHub: https://github.com/PolymathicAI/walrus
Website: https://polymathic-ai.org/blog/walrus/
Technical report: https://arxiv.org/abs/2511.15684Developer: Polymathic-AI
Released: November 2025
Parameters: 1.3B
Architecture: Transformer-based continuum-dynamics model with compute-adaptive patching, dimensional augmentation, and randomized compression
License: MIT
Modalities: Continuum-dynamics fields in, continuum-dynamics fields out; mixed 2D and 3D physical fields across acoustics, fluids, plasma, active matter, and astrophysics
Runs on: Desktop GPU
Formats: PyTorch, safetensors
On disk: 5.15GB PyTorch / 5.14GB safetensors -
PDEformer-2
GitHub: https://github.com/functoreality/pdeformer-2
Technical report: https://arxiv.org/abs/2507.15409Developer: Functoreality
Released: July 2025
Variants: Small, Fast, Base, Base-WDFE
Parameters: Small: 27.75M / Fast: 71.07M / Base: 82.65M / Base-WDFE: 84.70M
Architecture: Computational-graph encoder with a graph Transformer and an implicit neural representation decoder; Base-WDFE uses a weighted DeepSet function encoder for scattered-point inputs
License: Apache 2.0
Modalities: Symbolic 2D PDEs and numeric fields in, 2D solution fields out; arbitrary spatio-temporal query coordinates
Runs on: CPU
Formats: MindSpore checkpoint -
LANL MORPH
Hugging Face: https://huggingface.co/mahindrautela/MORPH
GitHub: https://github.com/lanl/MORPH
Technical report: https://arxiv.org/abs/2509.21670Developer: Los Alamos National Laboratory
Released: September 2025
Variants: Ti, S, M, L, XL
Architecture: Shape-agnostic vision transformer PDE surrogate with arbitrary data-modality support; full fine-tuning and LoRA adaptation
License: MIT
Modalities: PDE fields in, PDE fields out; mixed scalar and vector fields across 1D, 2D, and 3D systems
Runs on: Desktop GPU, NVIDIA only
Formats: PyTorch .pth
On disk: Ti: 110MB / S: 361MB / M: 1.36GB / L: 5.26GB / XL: 11.8GB -
General Physics Transformer
Hugging Face: https://huggingface.co/flwi/Physics-Foundation-Model
GitHub: https://github.com/FloWsnr/General-Physics-Transformer
Website: https://flowsnr.github.io/blog/physics-foundation-model/
Technical report: https://arxiv.org/abs/2509.13805Developer: FloWsnr
Released: September 2025
Variants: S, M, L, XL
Architecture: Transformer-based neural differentiator with numerical integration
License: MIT
Modalities: Physical fields in, physical fields out; eight fluid, solid, shock, thermal, and multiphase systems
Runs on: Desktop GPU, NVIDIA only
Formats: PyTorch .pth
On disk: S: 32.0MB / M: 446MB / L: 1.54GB / XL: 3.18GB -
LANL SPUS
Hugging Face: https://huggingface.co/siddik-lanl/spus-pde-unet-36m, https://huggingface.co/siddik-lanl/spus-pde-unet-36m-v2
GitHub: https://github.com/lanl/SPUS-Small-PDE-U-net-Solver
Technical report: https://arxiv.org/abs/2510.01370Developer: Los Alamos National Laboratory
Released: October 2025
Variants: v1, v2
Parameters: 36M
Resolution: 128x128 grid, 5 channels
Architecture: Residual U-Net with GELU, strided-convolution downsampling, transposed-convolution upsampling and skip connections; autoregressive next-step pretraining; v1 on four PDEgym compressible Euler datasets / v2 adds NS-Sines and NS-Gaussians, trained from fresh initialization
License: BSD-3-Clause
Modalities: 2D PDE fields in, 2D PDE fields out; density, horizontal velocity, vertical velocity, pressure, energy
Runs on: Laptop
Formats: safetensors
On disk: v1: 144.3MB / v2: 144.3MB safetensors -
JeongsLee MOTION
GitHub: https://github.com/JeongsLee/MOTION
Docs: https://github.com/JeongsLee/MOTION/blob/main/WEIGHTS.mdDeveloper: Jeongsu Lee, Kyung Hee University
Released: August 2026
Variants: MOTION, MOTION-S, 6-family joint-benchmark model, Poseidon-corpus IVP pretrain, six few-shot fine-tuned arms
Parameters: MOTION: 156.9M / MOTION-S: 20.7M / joint-benchmark: 114M / IVP pretrain and fine-tuned: 161.9M
Architecture: Per-step tendency as a gated sum of explicit physical-mechanism heads (transport, diffusion, pressure and density coupling, reaction, wave, curvature) plus an ungated state head; mechanism knockout and physical-equivalence tests; TensorFlow
License: Code: MIT / weights: CC-BY-4.0
Modalities: 2D and 3D PDE fields in, 2D and 3D PDE fields out; nineteen equation families across nine governing systems
Runs on: Inference from a released checkpoint: 1 GPU with 24GB or more, Linux
Formats: TensorFlow 2.15 named-array .npz in tar archives
On disk: MOTION: 580MB / MOTION-S: 152MB / joint-benchmark: 456MB / IVP pretrain: 648MB / IVP fine-tuned: 3.89GB tar
