When Latent Dynamics Meet Data Assimilation: A Benchmark of Koopman-Type Representations
Organizations: CEREA, ENPC, EDF R&D, Institut Polytechnique de Paris, Île-de-France, France · College of Computing and Data Science, Nanyang Technological University, Singapore · Dynamic Systems Lab, University College London, London, UK · Department of Mathematics, University of Wisconsin–Madison, Madison, WI, USA ·
Abstract
Koopman representations describe nonlinear dynamics through linear evolution in a transformed space, offering an attractive framework for their linear simplicity and generality. Yet these advantages do not necessarily translate into improved performance in deep latent data assimilation (LDA). We systematically compare seven representative methods and structural baselines across five systems spanning approximately periodic to strongly chaotic dynamics and a broad range of state dimensions. We evaluate forecast and assimilation accuracy, physical state recovery, latent-trajectory behavior, and dynamical consistency, and assess Gaussian error assumptions, objective convexity, and uniqueness of analysis solutions. Our results reveal a trade-off between representational flexibility and optimization tractability. Linear evolution of the latent state can stabilize short-term forecasts but restrict expressiveness in strongly chaotic and high-dimensional systems, without consistently improving assimilation accuracy. Parameterized linear operators offer greater flexibility, yet their joint optimization with the encoder--decoder remains nonconvex. Methods with closed-form analysis updates generally perform better and exhibit greater robustness in our experiments, although matrix inversion remains a computational bottleneck. These findings show that the practical benefits of Koopman structure depend on both sufficient observational information and the choice of assimilation procedure.
Figures & tables
| Model | Dynamics | Obs. coupling | Gauss. | Uncert. | Inference | Solution |
|---|---|---|---|---|---|---|
| DMD | Fixed linear | Latent residual | Yes | No | Convex 4D-Var | Closed form |
| NL-AE | Fixed nonlinear | Decoder | No | No | Nonlinear 4D-Var | Iterative |
| WL-AE | Fixed nonlinear | Decoder | No | No | Nonlinear 4D-Var | Iterative |
| KAE | Fixed linear | Decoder | No | No | Nonlinear 4D-Var | Iterative |
| Tensor-Var | Fixed linear | Latent residual ∗ | Yes | No | Convex 4D-Var | Closed form |
| DBF | Fixed linear | Inverse Obs. | Yes | Yes | Gaussian filtering | Closed form |
| Data set | State shape | Latent dimension | Pred. length | Assim. window | Spinup | Obs. every | Train | Val. | Test |
|---|---|---|---|---|---|---|---|---|---|
| Lorenz–96 | 40 | 5 | 2 | 1 | 128 | 16 | 16 | ||
| Cylinder | 512 | 32 | 6 | 4 | 2 | 39 | 5 | 5 | |
| Shallow water | 1152 | 40 | 6 | 4 | 2 | 560 | 168 | 168 | |
| ERA5–Coarse | 1024 | 14 | 3 | 2 | 1 | 1979–2015 | 2016 | 2018 | |
| ERA5–High | 5400 | 14 | 3 | 2 | 1 | 1979–2015 | 2016 | 2018 |
| Model | Lorenz–96 | Cylinder | Shallow water | ERA5–Coarse | ERA5–High |
|---|---|---|---|---|---|
| NL-AE | |||||
| WL-AE | |||||
| KAE | |||||
| Tensor-Var | |||||
| DMD | |||||
| DBF |
| Model | Lorenz–96 | Cylinder | Shallow water | ERA5–Coarse | ERA5–High |
|---|---|---|---|---|---|
| NL-AE | |||||
| WL-AE | |||||
| KAE | |||||
| Tensor-Var | |||||
| DMD | |||||
| DBF |
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
| Data set | Model | Parameters | Optimizer | Learning rate(s) | Batch | Epochs | Seeds |
|---|---|---|---|---|---|---|---|
| Lorenz–96 | DMD | – | – | – | – | – | – |
| Lorenz–96 | NL-AE | 380,440 | AdamW | 16 | 100 | 1, 2, 3 | |
| Lorenz–96 | WL-AE | 380,440 | AdamW | 16 | 100 | 1, 2, 3 | |
| Lorenz–96 | KAE | 380,640 | AdamW | 16 | 100 | 1, 2, 3 | |
| Lorenz–96 | Tensor-Var | 16,401,622 | AdamW | 64 | 60 | 1, 2, 3 | |
| Lorenz–96 | DBF | 549,520 | Adam | 16 | 100 | 1, 2, 3 |
| Method | Data set or training stage | Loss weights |
|---|---|---|
| NL-AE | All systems | ; |
| WL-AE | All systems | ; |
| KAE | All systems | ; ; for shallow water and for the other systems |
| Tensor-Var | State autoencoder | ; |
| Tensor-Var | Inverse-observation encoder | Latent state alignment weight: ; physical state reconstruction weight: |
| DMD | All systems | Not applicable |
| Data set | Short forecast lead | Long forecast lead |
|---|---|---|
| Lorenz–96 | 0.20 MTU | 0.40 MTU |
| Cylinder | 20 frames | 40 frames |
| Shallow water | 8 frames | 16 frames |
| ERA5–Coarse | 4 days | 8 days |
| ERA5–High | 4 days | 8 days |
| Data set | NL-AE | WL-AE | KAE | Tensor-Var | DMD |
|---|---|---|---|---|---|
| Lorenz–96 | |||||
| Cylinder | |||||
| Shallow water | |||||
| ERA5–Coarse | |||||
| ERA5–High |
| Parameter | Setting |
|---|---|
| Embedding dimension | 2 |
| Perplexity | |
| Initialization | PCA |
| Learning rate | auto |
| Early exaggeration | 12 |
| Maximum optimization iterations | 1,000 |
| Model | Metric | 5% | 15% | 30% | 50% |
|---|---|---|---|---|---|
| NL-AE | RRMSE | ||||
| SSIM | |||||
| WL-AE | RRMSE | ||||
| SSIM | |||||
| KAE | RRMSE | ||||
| SSIM |
| Model | Metric | 5% | 15% | 30% | 50% |
|---|---|---|---|---|---|
| NL-AE | RRMSE | ||||
| SSIM | |||||
| WL-AE | RRMSE | ||||
| SSIM | |||||
| KAE | RRMSE | ||||
| SSIM |
| Model | Metric | 5% | 15% | 30% | 50% |
|---|---|---|---|---|---|
| NL-AE | RRMSE | ||||
| SSIM | |||||
| WL-AE | RRMSE | ||||
| SSIM | |||||
| KAE | RRMSE | ||||
| SSIM |
| Model | Metric | 5% | 15% | 30% | 50% |
|---|---|---|---|---|---|
| NL-AE | RRMSE | ||||
| SSIM | |||||
| WL-AE | RRMSE | ||||
| SSIM | |||||
| KAE | RRMSE | ||||
| SSIM |
| Model | Metric | 5% | 15% | 30% | 50% |
|---|---|---|---|---|---|
| NL-AE | RRMSE | ||||
| SSIM | |||||
| WL-AE | RRMSE | ||||
| SSIM | |||||
| KAE | RRMSE | ||||
| SSIM |
Latent Dynamics Modeling
Latent dynamics modeling learns compact hidden-state representations and transition rules to predict, explain, and control evolving systems.
See all 271 papers