Physics-based dynamical models can successfully predict short-term sea ice trends, but they struggle to provide reliable subseasonal sea ice predictions over shorter time periods and are not flexible enough to provide real-time/quasi-real-time predictions. Therefore, we use deep learning to establish a purely remote sensing data-driven Arctic sea ice prediction system, IceFormer. IceFormer is trained by high-quality Arctic reanalysis data and remote sensing observations. It can provide daily sea ice concentration (SIC) and sea ice thickness (SIT) predictions for the pan-Arctic region with 45 days lead time. We show that IceFormer advances lightweight multivariate forecasting of Arctic sea ice. Driven by remote sensing data, the prediction performance of IceFormer is far superior to traditional baseline models and machine learning models. The root mean square error (RMSE) and spatial anomaly correlation coefficient (SACC) of SIC (averaged within 45 days) are 6.62% and 0.80, respectively. The RMSE and SACC of SIT (averaged within 45 days) are 0.04m and 0.93, respectively. IceFormer can successfully capture the sea ice evolution on a seasonal scale (especially in the melting season) and the monthly average SACC is stable at around 0.70. After comparison, the SIC prediction performance of IceFormer is better than that of a Coupled Ensemble Numerical Forecast Model (SEAS5) from May to November. The information transfer analysis shows that the simultaneous prediction of SIC and SIT can effectively improve the predictability and explainability of sea ice.

We explore to what extent data-driven prediction models have skills in forecasting daily sea-surface temperature (SST), which are comparable to or perform better than current physics-based operational systems over long-range forecast horizons. Three hybrid deep learning-based models are developed within the South China Sea (SCS) basin by integrating deep neural networks (back propagation, long short-term memory, and gated recurrent unit) with traditional empirical orthogonal function analysis and empirical mode decomposition. Utilizing a 40-year (1982–2021) satellite-based daily SST time series on a 0.25° grid, we train these models on the first 32 years (1982–2013) of detrended SST anomaly (SSTA) data. Their predictive accuracies are then validated using data from 2014 and tested over the subsequent seven years (2015–2021). The models’ forecast skills are assessed using spatial anomaly correlation coefficient (ACC) and root-mean-square error (RMSE), with ACC proving to be a stricter metric. A forecast skill horizon, defined as the lead time before ACC drops below 0.6, is determined to be 50 days. The models are equally capable of achieving a basin-wide average ACC of ~0.62 and an RMSE of ~0.48 °C at this horizon, indicating a 36% improvement in RMSE over climatology. This implies that on average the forecast skill horizon for these models is beyond the available forecast length. Analysis of one model, the BP neural network, reveals a variable forecast skill horizon (5 to 50 days) for each individual day, showing that it can adapt to different time scales. This adaptability seems to be influenced by a number of mechanisms arising from the evident regional and global atmosphere–ocean coupling variations on time scales ranging from intraseasonal to decadal in the SSTA of the SCS basin.

On ensemble‐based coupled data assimilation, cross‐component parameter estimation (CPE), has not been as extensively developed and applied as weakly coupled state and parameter estimation along with cross‐component state estimation. This discrepancy is partially attributed to the lack of emphasis on the instantaneous response of coupled model states with respect to parameters across different components. We define so‐called response as the instantaneous parameter sensitivity (IPS). Under the framework of sequential assimilation, the prior information heavily relies on the IPS of coupled states with different time scales. Based on the IPS analysis for an intermediate coupled model, a series of twin experiments of state and parameter estimation are conducted, in which an IPS‐inspired adaptive inflation scheme for parameter ensemble is introduced. Results show that the success of a parameter estimation strategy is closely tied to the significant IPS of the observed state to the parameter targeted for optimization, as it maintains a high signal‐to‐noise ratio in the error covariance between parameter and prior state, thereby enhancing parameter estimation. An interesting finding in the context of IPS‐based CPE is: an atmospheric parameter can be successfully estimated by assimilating observations from slow‐varying oceanic component, but not vice versa. In comparison with cross‐component state estimation, successful CPE significantly enhances the estimation accuracy of coupled states by mitigating model bias.

This study introduces two distinct post-processing strategies to address systematic biases in sea surface temperature (SST) numerical forecasts, thereby enhancing SST predictive accuracy. The first strategy implements a spatiotemporal four-dimensional multi-grid analysis (4D-MGA) scheme within a three-dimensional variational (3D-Var) data assimilation framework. The second strategy establishes a hybrid deep learning architecture integrating empirical orthogonal function (EOF) analysis, empirical mode decomposition (EMD), and a backpropagation (BP) neural network (designated as EE–BP). The 4D-MGA strategy dynamically corrects systematic biases through a temporally coherent extrapolation of analysis increments, leveraging its inherent capability to characterize intrinsic temporal correlations in model error evolution. In contrast, the EE–BP strategy develops a bias correction model by learning the systematic biases of the SST numerical forecasts. Utilizing a satellite fusion SST dataset, this study conducted bias correction experiments that specifically addressed the daily SST numerical forecasts with 7-day lead times in the Kuroshio region south of Japan during 2017, systematically quantifying the respective error reduction potentials of both strategies. Quantitative verification reveals that EE–BP delivers enhanced predictive skill across all forecast horizons, achieving 18.1–22.7% root–mean–square error reduction compared to 1.2–9.1% attained by 4D-MGA. This demonstrates deep learning’s unique advantage in capturing nonlinear bias evolution patterns.

Advances in data assimilation (DA) methods have greatly improved the accuracy of Earth system predictions. To fuse multi-source data and reconstruct the nonlinear evolution missing from observations, geoscientists are developing future-oriented DA methods. In this paper, we redesign a purely data-driven latent space DA framework (DeepDA) that employs a generative artificial intelligence model to capture the nonlinear evolution in sea surface temperature. Under variational constraints, DeepDA embedded with nonlinear features can effectively fuse heterogeneous data. The results show that DeepDA remains highly stable in capturing and generating nonlinear evolutions even when a large amount of observational information is missing. It can be found that when only 10% of the observation information is available, the error increase of DeepDA does not exceed 40%. Furthermore, DeepDA has been shown to be robust in the fusion of real observations and ensemble simulations. In particular, this paper provides a mechanism analysis of the nonlinear evolution generated by DeepDA from the perspective of physical patterns, which reveals the inherent explainability of our DL model in capturing multi-scale ocean signals.