Reproduce “Online machine-learning forecast uncertainty estimation for sequential data assimilation” on OpenResearch

About this paper

Maximiliano A. Sacco, Manuel Pulido, Juan J. Ruiz, Pierre Tandeo

The UnnKF system integrates a convolutional neural network into a Kalman-like data assimilation framework to estimate state-dependent forecast error covariance using a single model integration. This approach achieves analysis RMSE comparable to a 100-member Ensemble Kalman Filter in a Lorenz'96 model while significantly outperforming a 5-member EnKF by effectively filtering sampling noise in training data.

Key insights

• Machine learning models can effectively learn and filter sampling noise from limited-ensemble training data, leading to more robust and accurate forecast error covariance estimates than direct ensemble samples.
• A custom 'extended-MSE' (eMSE) loss function provides numerical stability for training neural networks to estimate localized, state-dependent forecast error covariance matrices in high-dimensional data assimilation contexts.
• The performance of the ML-based covariance estimator is sensitive to the choice of forecast error proxy used for training, with approximations closer to the true error yielding better analysis quality.

Summary by alphaXiv.