Benchmarking Stochastic Interpolants for Modeling Physical Systems
Organizations: Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Department of Geophysical Sciences, Committee on Computational and Applied Mathematics, University of Chicago, Chicago, IL 60637, USA · LMD/IPSL, CNRS, ENS, Université PSL, École Polytechnique, Institut Polytechnique de Paris, Sorbonne Université, Réseau de Transport d’Électricité (RTE), Paris, France
Abstract
Generative models have recently emerged as powerful surrogates for physical systems, demonstrating increased accuracy, stability, and/or statistical fidelity. However, most approaches rely on iteratively denoising a Gaussian, a choice that may not be the most effective for autoregressive prediction tasks, where the current and future states are often closely related. We therefore benchmark generative models based on stochastic interpolants, which can learn maps between arbitrary densities, against standard Gaussian-denoising frameworks across a range of fluid and climate systems. Following prior work, we consider a transport map for dynamical systems that learns an interpolant between a point mass at the current state and a conditional distribution of future states. We find that stochastic interpolants can match or outperform Gaussian-denoising baselines in pointwise accuracy, spectral fidelity, and probabilistic skill. Furthermore, we identify a tradeoff where deterministic sampling can improve pointwise accuracy but smooth predictions, whereas stochastic sampling recovers spectral content and calibration but requires more function evaluations, as well as ensemble averaging to recover pointwise accuracy. As a whole, these results establish stochastic interpolants as a competitive baseline for generative physical emulation, and also indicate that the training loss, sampler, and transport map should be chosen with the physical system and target metric in mind.
Figures & tables
| Training Objective | Sampling Procedure | |
|---|---|---|
| DDPM | ||
| DDIM | ||
| EDM | (Euler) | |
| TSM | (1-step) | |
| FM | (Euler) | |
| SI | (Euler-Maruyama) |
| Surface Variables (8) | Atmospheric Variables (5) | Forcing Variables (6) | Pressure Levels (13) |
|---|---|---|---|
| lwe of water evaporation ( evap ) surface runoff ( mrro ) lwe of soil moisture content ( mrso ) log surface pressure ( pl ) 12h acc. precipitation ( pr_12h ) 6h acc. precipitation ( pr_6h ) air temperature 2m ( t2m ) surface temperature ( ts ) | specific humidity ( hus ) air temperature ( ta ) eastward wind ( ua ) northward wind ( va ) geopotential ( zg ) | land sea mask ( lsm ) surface geopotential ( sg ) surface roughness length ( z0 ) TOA Incident Radiation ( rsdt ) sea ice cover ( sic ) sea surface temperature ( sst ) | 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 |
| Model | NFE | VRMSE | SRMSE L | SRMSE M | SRMSE H | SSR | CRPS | Ens. VRMSE |
|---|---|---|---|---|---|---|---|---|
| FNO | 1 | 0.589 0.009 | 0.079 0.009 | 0.046 0.003 | 0.406 0.013 | — | 1.507 0.028 | 0.589 0.009 |
| DDPM | 100 | 0.539 0.002 | 0.125 0.020 | 0.076 0.004 | 0.795 0.000 | 0.937 0.018 | 0.707 0.008 | 0.413 0.003 |
| DDIM | 10 | 0.604 0.004 | 0.275 0.020 | 0.170 0.002 | 0.801 0.001 | 0.355 0.007 | 1.180 0.009 | 0.557 0.003 |
| TSM | 10 | 1.914 1.088 | 0.399 0.072 | 0.463 0.151 | 0.891 0.039 | 0.748 0.144 | 3.343 2.206 | 1.576 0.892 |
| EDM (Det.) | 10 | 0.426 0.002 | 0.100 0.005 | 0.064 0.002 | 0.793 0.000 | 0.744 0.025 | 0.583 0.013 | 0.340 0.005 |
| EDM (Sto.) | 20 | 0.427 0.001 | 0.088 0.013 | 0.061 0.002 | 0.793 0.001 | 0.720 0.016 | 0.598 0.004 | 0.346 0.004 |
| Model | NFE | VRMSE | SRMSE L | SRMSE M | SRMSE H | SSR | CRPS | Ens. VRMSE |
|---|---|---|---|---|---|---|---|---|
| FNO | 1 | 0.589 0.065 | 0.176 0.029 | 3.103 0.412 | 2.783 0.261 | — | 0.097 0.013 | 0.589 0.065 |
| DDPM | 100 | 0.468 0.009 | 0.147 0.009 | 0.107 0.003 | 1.535 0.027 | 1.016 0.018 | 0.036 0.000 | 0.349 0.004 |
| DDIM | 10 | 0.452 0.002 | 0.523 0.010 | 0.713 0.033 | 2.985 0.047 | 0.590 0.008 | 0.049 0.000 | 0.396 0.003 |
| TSM | 10 | 0.605 0.045 | 0.348 0.042 | 0.391 0.057 | 3.518 0.165 | 0.577 0.033 | 0.072 0.007 | 0.538 0.037 |
| EDM (Det.) | 10 | 0.415 0.008 | 0.146 0.003 | 0.167 0.007 | 2.241 0.024 | 0.840 0.002 | 0.034 0.000 | 0.329 0.002 |
| EDM (Sto.) | 50 | 0.430 0.005 | 0.130 0.006 | 0.116 0.003 | 1.813 0.014 | 0.921 0.009 | 0.034 0.000 | 0.329 0.001 |
| Var. | Day | SFNO | DDPM | DDIM | EDM (Det.) | EDM (Sto.) | FM | SI (Det.) | SI (Sto.) |
|---|---|---|---|---|---|---|---|---|---|
| NFE | 1 | 100 | 10 | 10 | 20 | 10 | 10 | 20 | |
| [m] | 1 | 7.8 0.05 | 6.5 0.09 | 6.4 0.06 | 6.0 0.01 | 6.0 0.01 | 6.1 0.01 | 5.9 0.02 | 6.0 0.04 |
| 3 | 20.6 0.09 | 13.2 0.23 | 13.3 0.17 | 11.8 0.15 | 11.8 0.11 | 12.0 0.14 | 11.5 0.12 | 11.5 0.11 | |
| 5 | 35.6 0.23 | 22.4 0.33 | 22.7 0.32 | 20.4 0.21 | 20.3 0.19 | 20.8 0.32 | 20.0 0.09 | 19.7 0.08 | |
| 10 | 70.7 1.04 | 46.1 0.35 | 46.5 0.18 | 44.1 0.48 | 43.8 0.25 | 45.2 0.28 | 49.3 0.05 | 42.9 0.59 | |
| [K] | 1 | 1.02 0.007 | 0.74 0.003 | 0.75 0.004 | 0.71 0.000 | 0.71 0.003 | 0.71 0.002 | 0.70 0.002 | 0.70 0.001 |
| Var: | [m] | [K] | [K] | [m/s] | [mm] | [g/kg] |
|---|---|---|---|---|---|---|
| DDIM | 9.77 1.04 | 0.53 0.14 | 0.45 0.13 | 1.33 0.15 | 0.106 0.001 | 0.426 0.015 |
| EDM (Sto.) | 4.31 0.71 | 0.34 0.01 | 0.28 0.01 | 0.63 0.11 | 0.069 0.006 | 0.270 0.009 |
| FM | 4.18 0.71 | 0.28 0.01 | 0.25 0.02 | 0.55 0.06 | 0.062 0.002 | 0.202 0.017 |
| SI (Sto.) | 3.95 0.30 | 0.25 0.01 | 0.23 0.00 | 0.49 0.02 | 0.048 0.001 | 0.128 0.006 |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Metric | FM pixel | FM latent | SI pixel | SI latent |
|---|---|---|---|---|
| NRMSE | 0.812 | 0.649 | 0.966 | 0.609 |
| AE | FM | SI | |
|---|---|---|---|
| 64 | 0.0252 | 0.619 | 0.605 |
| 256 | 0.0457 | 0.649 | 0.622 |
| : | 0.1 | 0.5 | 1 | 3 |
|---|---|---|---|---|
| NRMSE | 0.702 | 0.632 | 0.609 | 0.803 |
| Model | Schedule | Scale | VRMSE | SRMSE low | SRMSE mid | SRMSE high |
|---|---|---|---|---|---|---|
| DDPM | Linear | 0.691 | 0.137 | 0.074 | 0.794 | |
| DDIM | Linear | 0.797 | 1.840 | 0.831 | 0.977 | |
| TSM | Linear | 0.969 | 1.413 | 2.082 | 0.932 | |
| DDPM | Linear | 0.655 | 0.085 | 0.056 | 0.794 | |
| DDIM | Linear | 0.744 | 0.284 | 0.208 | 0.804 | |
| TSM | Linear | 0.827 | 1.541 | 0.892 | 0.896 |
| Sampler | VRMSE | SRMSE low | SRMSE mid | SRMSE high | ||
|---|---|---|---|---|---|---|
| ODE | 0.616 | 0.084 | 0.062 | 0.793 | ||
| ODE | 0.682 | 0.236 | 0.061 | 0.792 | ||
| ODE | 0.692 | 0.252 | 0.074 | 0.792 | ||
| ODE | 0.606 | 0.073 | 0.047 | 0.792 | ||
| ODE | 0.685 | 0.194 | 0.059 | 0.791 | ||
| ODE | 0.688 | 0.241 | 0.069 | 0.792 |
| Parameter | Kolmogorov Flow | Rayleigh-Benard | PlaSim |
|---|---|---|---|
| Autoencoder | |||
| Input resolution | |||
| Latent resolution | |||
| Downsampling factor | 64 | 64 | 32 |
| Channel widths | [64, 128, 256, 256] | [64, 128, 256, 512] | [128, 256, 512] |
| ResBlocks per stage | 3 | 4 | 2 |
| Model | NFE | VRMSE | SRMSE L | SRMSE M | SRMSE H | SSR | CRPS | Ens. VRMSE |
|---|---|---|---|---|---|---|---|---|
| (100-step rollout) | ||||||||
| FNO | 1 | 0.591 | 0.070 | 0.043 | 0.398 | — | 1.522 | 0.591 |
| DDPM | 100 | 0.539 | 0.122 | 0.079 | 0.795 | 0.952 | 0.702 | 0.411 |
| DDIM | 10 | 0.605 | 0.260 | 0.171 | 0.802 | 0.363 | 1.172 | 0.554 |
| TSM | 10 | 1.004 | 0.411 | 0.336 | 0.854 | 0.591 | 1.603 | 0.845 |
| EDM (Det.) | 10 | 0.423 | 0.099 | 0.065 | 0.793 | 0.771 | 0.574 | 0.337 |
Stochastic Interpolants
Stochastic interpolants learn continuous probability paths that transform noise or one distribution into another.
See all 26 papers