Multi-field sea ice data assimilation with a variational autoencoder

ICLR 2026 Conference Submission18876 Authors

Keywords: Machine Learning, Deep learning, Data assim-ilation, NEMO, Ice concentration, Neural network, Variational Autoencoder, VAE, 3D-VAR

Abstract:

We introduce an assimilation approach that simultaneously processes several physical fields, leveraging a modern variational autoencoder (VAE) architecture enhanced with pixel-wise self-attention mechanism to capture complex spatial and cross-field correlations. We test our approach in assimilation of the real-world satellite observations (Sentinel-3 SRAL and AMSR2) in simulation data from the NEMO ocean model with integrated sea ice component (SI3). We demonstrate that our approach effectively works with sparse and noisy observations, reducing errors in sea ice concentration estimates and improving simulation accuracy. Also we integrate our neural assimilation solution in the NEMO restart mechanism and evaluate its influence on the operational forecasts. This work bridges the gap between machine learning-based assimilation and practical ocean modeling, offering a scalable non-Gaussian alternative to traditional methods like 3D-VAR.
and improving forecast accuracy. Crucially, we demonstrate the compatibility of the neural assimilation solution with the NEMO restart mechanism, enabling seamless integration into operational forecasting pipelines. This work bridges the gap between machine learning-based assimilation and practical ocean modeling, offering a scalable, non-Gaussian alternative to traditional methods like 3D-VAR.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 18876