Physics-Knowledge-Guided Hybrid Neural Learning for Arctic Sea Ice Concentration Evolution and Short-Range Prediction

Deep Learning Benchmarks for short-term Arctic Sea Ice Forecasting

Abstract. The rapid decline of Arctic sea ice has increased spatiotemporal variability in the core marginal seas along the Northern Sea Route (NSR), thereby hindering reliable short-term forecasting. While deep learning models offer computationally efficient alternatives to physics-based numerical models, previous studies have often relied on global average errors and have provided limited assessment of boundary-preservation performance. In addition, few studies have systematically compared spatial feature extraction strategies within non-recurrent spatiotemporal architectures for sea ice forecasting. To address this gap, this study benchmarks CNN backbone and Transformer backbone families for 10-day sea ice forecasting. The evaluation focuses on five dynamic marginal seas along the NSR. Metrics included Integrated Ice Edge Error (IIEE), Mean Boundary Error (MBE), Intersection over Union (IoU), and Anomaly Correlation Coefficient (ACC). Results indicate that CNN-based models, including PoolFormer, generally show more favorable performance in boundary preservation and prediction stability than Transformer-based models. Regionally, the Barents and Kara Seas were more difficult to predict, whereas the Laptev and East Siberian Seas were relatively more predictable. The Chukchi Sea exhibited particularly high uncertainty during rapid summer ice retreat. Across all models, boundary-based performance degraded significantly as lead time increased, with MBE exceeding 30 km from T+5 onward. These results suggest that, under the current univariate setting, boundary-preservation performance remains relatively stable at short lead times, but degrades thereafter. Overall, CNN-based non-recurrent architectures show relative strengths over the NSR test regions, although future work incorporating multivariate inputs is needed to extend reliable lead times.

Physics-informed machine learning for south pacific sea-level rise: integrating ocean-atmosphere-land interactions

Deep Learning-Based Statistical Downscaling of Sea Surface Temperature Using a Residual Corrective Neural Network

This study proposes a novel deep learning framework that uses a U-Net to generate an initial high-resolution SST estimate, which is subsequently refined using a residual corrective approach, and progressively refines initial U-Net predictions by incorporating dynamically scaled residuals at each step, enabling accurate capture of broad patterns and fine-grained features such as eddies and fronts.

DLESyM-Ocean: A Deep Learning Probabilistic Global Model for Simulating Present-Day Upper Ocean and Sea Ice

DLESyM-Ocean is stable when autoregressively run for multi-year simulations and produces a climatology and variability with minimal bias compared with reanalysis, and the computational efficiency of DLESyM-Ocean makes it a promising tool for subseasonal to seasonal forecasting.

Literature Review of Hybrid Neural–Physics Flood-Relevant Streamflow Forecasting with Cross-Event and Cross-Basin Generalization

Reliable flood-relevant streamflow forecasting remains difficult at operational scale because physically based forecasting systems provide national coverage but cannot be directly verified at the overwhelming majority of ungauged river reaches. This study reconstructs and strengthens a hybrid neural–physics framework in which the National Water Model (NWM) supplies the physically constrained baseline forecast and an attribute-conditioned deep neural network learns forecast-error magnitude from gauged basins, then transfers that knowledge to ungauged basins. The emphasis is not on replacing process-based hydrology with a black-box predictor, but on improving forecast reliability under cross-event and cross-basin generalization. The empirical setting consists of 979 archived daily long-range forecasts, 30 forecast lead days, 16-member NWM ensembles, and 389 quality-controlled U.S. Geological Survey gauging stations across Alabama and Georgia, complemented by static watershed descriptors and gridded soil-moisture diagnostics. Exploratory analysis shows that forecast skill is strongly conditioned by basin properties: larger and forested watersheds are predicted more accurately than smaller and urban watersheds, while urban basins show severe positive bias, weaker anomaly correlation, and degraded hydrograph timing. These diagnostics motivate an error-learning postprocessor that maps NWM forecast magnitude and basin attributes to a non-negative error estimate and converts the deterministic physics forecast into a corrected predictive interval. Under a temporally disjoint evaluation intended to test cross-event transfer, observation coverage increases from 0.21±0.01 for the raw NWM ensemble range to 0.82±0.03 for the hybrid model; a repeated spatial holdout yields comparable performance (0.82±0.01). The gains are largest where the underlying physics model is weakest, especially in developed watersheds. The main limitation is that improved reliability is achieved through wider predictive intervals rather than sharper hydrographs. The study therefore supports a defensible conclusion of hybrid neural–physics postprocessing can materially improve flood-relevant streamflow reliability at ungauged basins, but interval sharpness and external cross-region validation remain open research needs.

Deep learning for multi-step forecasting of stratospheric wind and temperature profiles over the Tibetan plateau

The MTPV-HDRNet model is proposed as a fixed-site, multi-level intelligent forecasting model driven by ECMWF ERA5 pressure-level reanalysis data, adopting a hybrid encoder–decoder architecture that explicitly represents temporal evolution and vertical stratification in a decoupled but complementary manner.