Research map / Probabilistic and scientific modeling / Differential equations and PDEs
P3D: Highly Scalable 3D Neural Surrogates for Physics Simulations with Global Context
Benjamin Holzschuh, Georg Kohl, Florian Redinger, Nils Thuerey
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add transformer backbone on top of conv encoder/decoder (ablation) | P3D-S · joint PDE, crop 64^3 | validation MSE x1e-3: 5.40 -> 2.77 | Improved |
| replace patch tokenizer with large convolutional encoder/decoder (ablation) | P3D-S · joint PDE, crop 64^3 | validation MSE x1e-3: 3.78 -> 2.77 | Improved |
| add PDE type conditioning (ablation) | P3D-S · joint PDE, crop 64^3 | validation MSE x1e-3: 2.94 -> 2.77 | Improved |
| enable shifted windows (ablation) | P3D-S · joint PDE | validation MSE x1e-3: 2.77 -> 2.84 (64^3) | No clear effect |
| increase window size from 4 to 8 (ablation) | P3D-S · joint PDE | validation MSE x1e-3: 2.77 -> 2.90 (64^3), 1.35 -> 1.32 (128^3) | No clear effect |
| scale model size from S to L | P3D · joint PDE, crop 128^3 | nRMSE x1e-2: 3.33 -> 2.08 | Improved |
| train on larger crops | P3D-S · joint PDE | nRMSE x1e-2: 6.27 (32^3) -> 3.76 (64^3) -> 3.33 (128^3) | Improved |
| run full-domain inference via translation equivariance instead of independent crops | P3D-S · isotropic turbulence 512^3 | RMSE x1e-2: 1.90 -> 1.60 | Improved |
| add context network with finetuning to crop-pretrained model (ablation) | P3D-L · turbulent channel flow | mean L2 x1e-5: 5862.81 -> 23.6 | Improved |
| add region tokens in finetuning (ablation) | P3D-L · turbulent channel flow | mean L2 x1e-5: 4541 -> 23.6 | Improved |
| finetune decoder only for longer (20 to 500 epochs) (ablation) | P3D-L · turbulent channel flow | mean L2 x1e-5: 941 -> 16.8 | Improved |
| train P3D on crops and scale to full domain vs train on full domain | P3D-L · turbulent channel flow | mean L2 x1e-5: 3.02 (full domain) vs 16.8 (decoder-only finetune) | Hurt |
Closest papers by meaning
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.89)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.88)
- PEINR: A Physics-enhanced Implicit Neural Representation for High-Fidelity Flow Field Reconstruction (ICML 2025 · similarity 0.87)
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (NeurIPS 2024 · similarity 0.87)
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs (NeurIPS 2025 · similarity 0.87)
- EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
