arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Neural CGM unites closed‑form data assimilation with neural dynamics for efficient state estimation | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Machine LearningEnglishPublished

Neural CGM unites closed‑form data assimilation with neural dynamics for efficient state estimation

November 22, 2026arXiv: 2611.26463v1

Researchers have built a data‑driven model, the Neural Conditional Gaussian Model (Neural CGM), that learns a latent‑space conditional Gaussian dynamics. A nonlinear encoder maps the observed and hidden physical states into this latent space, while the dynamics follow a closed‑form Gaussian structure. In the latent space, data‑assimilation updates can be calculated analytically, and a distribution‑aware decoder maps the posterior mean and covariance back to the physical variables.

On the Lorenz‑63 benchmark, Neural CGM forecast a root‑mean‑square error (RMSE) of 1.6907 and a data‑assimilation RMSE of 1.3882. Its negative log‑likelihood (NLL) was 3.8068 in the data‑assimilation step, and the entire update took 0.8772 s. In comparison, the previous best data‑driven method, CGKN, gave a forecast RMSE of 1.6941, DA RMSE of 1.4174, DA NLL of 3.9126, and DA time of 0.9367 s.

On the larger Lorenz‑96 system, Neural CGM achieved a forecast RMSE of 0.7353, DA RMSE of 3.1927, DA NLL of 92.1253, and DA time of 3.5831 s, slightly outperforming CGKN which reached 0.7356 forecast RMSE, 3.1962 DA RMSE, 95.2390 DA NLL, and 2.8629 s.

For the chaotic Kuramoto‑Sivashinsky equation, Neural CGM produced a forecast RMSE of 0.0554, DA RMSE of 0.3846, DA NLL of 54.6653, and DA time of 1.8307 s, better than CGKN's 0.0621 forecast RMSE, 0.4222 DA RMSE, 166.1100 DA NLL, and 1.3559 s.

The model’s training requires paired trajectories of both observed and unobserved variables. If such data are unavailable, the method cannot be applied without further development. The current likelihood calculation uses only diagonal physical‑space posterior covariances.

Thus, Neural CGM represents the first approach that combines closed‑form latent‑space data assimilation, a distribution‑aware decoding step, and likelihood‑based training to deliver accurate state estimation, forecasting, and uncertainty quantification for partially observed nonlinear systems.

Source paper

Neural Conditional Gaussian Model: Efficient Data Assimilation with Sequential Uncertainty Quantification for Nonlinear Dynamics

Machine LearningMathematics
View on arXiv