Ultra Fast PDE Solving via Physics Guided Few-step Diffusion
Organizations: Purdue University · hi-Lab, Xiaohongshu Inc
Abstract
Diffusion-based models have demonstrated impressive accuracy and generalization in solving partial differential equations (PDEs). However, their practical utility still suffers from high sampling costs and insufficient physical consistency, mainly due to many-step iterative sampling and insufficient physics guidance. To address these issues, we propose Phys-Instruct, a physics-guided distillation framework that compresses a pre-trained diffusion PDE solver into a few-step generator while explicitly enforcing PDE consistency through distillation guidance. We reformulate the training objective as a constrained problem with PDE residual, and derive a score-function identity that leads to a practical physics-guided generator update. Across five PDE benchmarks, Phys-Instruct achieves orders-of-magnitude faster inference while reducing the PDE Error by factors of to , compared with the corresponding 200-step EDM teachers. Moreover, the resulting unconditional student model functions as a compact prior, enabling efficient and physically consistent inference for various downstream conditional tasks. Our results indicate that Phys-Instruct is an effective and efficient framework for ultra-fast PDE solving powered by generative models.
Figures & tables
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Step | Inference Latency (s) | PDE Error | SWD | Recall | Diversity/T200 |
|---|---|---|---|---|---|---|
| Phys-Instruct | 1 | 0.012 | 0.171 | 0.071 | 0.674 | 0.658 |
| 2 | 0.036 | 0.118 | 0.060 | 0.830 | 0.896 | |
| 4 | 0.084 | 0.055 | 0.049 | 0.930 | 0.890 | |
| 8 | 0.225 | 0.046 | 0.046 | 0.919 | 0.905 | |
| EDM | 200 | 4.919 | 0.860 | 0.043 | 0.980 | 1.000 |
| Benchmark | Teacher | Distill | D/T |
|---|---|---|---|
| Darcy Flow | 689.4 | 506.9 | 0.735 |
| Poisson | 689.4 | 506.9 | 0.735 |
| 1609.0 | 1191.1 | 0.740 | |
| Burgers’ | — | 3185.3 | — |
| Helmholtz | — | 3364.9 | — |
| Benchmark | PDE | PDE/G | Total | ||
|---|---|---|---|---|---|
| Darcy | 0.086 | 0.128 | 0.042 | 33.2% | 0.213 |
| Poisson | 0.086 | 0.128 | 0.042 | 33.2% | 0.213 |
| 0.269 | 0.440 | 0.141 | 32.0% | 0.709 | |
| Burgers’ | 1.328 | 1.990 | 0.663 | 33.3% | 3.318 |
| Helmholtz | 1.328 | 1.991 | 0.663 | 33.3% | 3.319 |
| Benchmark | Resolution | Step | PDE Error |
|---|---|---|---|
| Darcy Flow | 1 | ||
| 2 | |||
| 3 | |||
| 4 | |||
| Poisson | 1 | ||
| 2 |
| Resolution | Step | PDE Error | Ratio to Teacher | Time (ms) | Peak Mem (MB) |
|---|---|---|---|---|---|
| 1 | 0.02 | 69% | 16.87 | 357.37 | |
| 1 | 0.52 | 41% | 19.38 | 452.95 | |
| 1 | 3.52 | 59% | 62.59 | 894.08 | |
| 1 | 12.33 | 34% | 317.00 | 3075.57 |
| Method | Step | FD4 Residual | Re-solve | Re-solve |
|---|---|---|---|---|
| Phys-Instruct | 1 | |||
| 4 | ||||
| EDM | 200 |
| Method | Step | Spectral Residual | Mass Defect | Energy-Balance Defect |
|---|---|---|---|---|
| Phys-Instruct | 1 | |||
| 4 | ||||
| EDM | 200 |
| Method | Step | PDE Error | Spectral PDE Error | Spectrum Mismatch |
|---|---|---|---|---|
| EDM | 8 | |||
| Phys-Instruct | 8 | |||
| EDM | 200 | (reference) |
| Status | |
|---|---|
| Unstable (early failure; PDE error remains high) | |
| Stable, but weak improvement | |
| Stable | |
| Stable | |
| Stable | |
| Diverged |
| PDE Error ( ) | |
|---|---|
| 1.00 | |
| 1.50 (default) | |
| 3.00 |
| PDE Error ( ) | |||
|---|---|---|---|
| Method | Step | Euler | Heun |
| Phys-Instruct | 1 | ||
| 2 | |||
| 3 | |||
| 4 | |||
| Step | Darcy Flow | Helmholtz | |||
|---|---|---|---|---|---|
| Forward | Inverse | Forward | Inverse | ||
| Phys-Instruct (ours) | 1 | 0.50% | 0.35% | 1.86% | 4.58% |
| Transolver | – | 2.5% | 7.5% | 2.3% | 12.0% |
| FNO | – | 5.3% | 5.6% | 11.1% | 5.0% |
| PINO | – | 4.0% | 2.1% | 4.9% | 4.9% |
| DeepONet | – | 12.3% | 8.4% | 17.8% | 28.1% |
| Darcy Flow | Poisson | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Method | Step | Forward | Inverse | Forward | Inverse | ||||
| Rel. Err. | PDE Err. | Abs. Err. | PDE Err. | Rel. Err. | PDE Err. | Rel. Err. | PDE Err. | ||
| Phys-Instruct | 1 | 0.29% | 0.28 | 0.49% | 0.22 | 2.49% | 0.10 | 8.64% | 0.015 |
| DiffusionPDE | 2000 | 4.61% | 0.53 | 7.71% | 1.68 | 7.48% | 0.22 | 13.34% | 0.091 |
| CoCoGen | 2000 | 4.39% | 0.45 | 7.44% | 1.59 | 7.25% | 0.20 | 13.17% | 0.071 |
| PIDM | 2000 | 7.34% | 0.34 | 9.83% | 1.76 | 5.28% | 0.12 | 17.48% | 0.077 |
| Partial Observation | |||
|---|---|---|---|
| Method | Step | Rel. Error | PDE Error ( ) |
| Phys-Instruct | 1 | ||
| DiffusionPDE | 200 | ||
| CoCoGen | 200 | ||
| PIDM | 200 | ||
| DiffusionPDE | 2000 | ||
Neural PDE Solvers
Neural PDE solvers use trainable models to approximate solutions or solution operators for partial differential equations, often with physical constraints.
See all 297 papers