Review of “ MET-AICE v1.0: an operational data-driven sea ice prediction system for the European Arctic” by Palerme et al. Submitted to Geoscientific Model Development.

General comments

MET-AICE v1.0 is the first operational, data-driven sea ice prediction system specifically designed for short-term forecasts (1-10 days) in the European Arctic. The system is optimised for operational utility and higher spatial resolution, making it suitable for day-to-day maritime applications. The development of the MET-AICE system is particularly timely given the increasing demand for reliable, short-term, high-resolution sea ice forecasts, driven by increased maritime activity and heightened navigational risks associated with changing sea ice cover. 

MET-AICE was trained on weekly AMSR2 weekly sea ice concentration data at 5-km resolution 2020 from the recently published reSICCI3LF algorithm, covering the period from 2013 to 2020. During training, the neural network models were iteratively updated over 100 epochs to minimize the mean squared error between the predicted SIC and the AMSR2 SIC observations. The system incorporates several predictors, including 9-km resolution ECMWF weather forecasts (2-m temperature and 10-m wind components), AMSR2 SIC observations from the day preceding the forecast start date, and a land-sea mask. MET-AICE uses a convolutional neural network with a U-Net architecture, designed specifically to capture spatial hierarchies in the input data. Operational forecasts have been generated since March 2024, with validation described in the manuscript covering a year-long period from April 2024 to March 2025. Despite demonstrated strengths in computational efficiency and accuracy compared to the Barents-2.5 km EPS model and other validation datasets, MET-AICE experiences reduced accuracy in coastal regions and diminished predictive skill during sea ice minimum periods, primarily related to inherent limitations in the input datasets. The current version of MET-AICE provides deterministic forecasts of sea ice concentration, which become smoother as the lead time increases. In future iterations, the authors plan to incorporate ensemble and probabilistic approaches to better quantify and represent the forecast uncertainty. 

The paper is generally well written and structured, providing an important contribution towards operational high-resolution sea ice forecasting. However, several points need clarification before I can recommend the manuscript for publication. 

1. I found the model description quite hard to follow. I wonder if you could include a flow diagram that shows the data inputs and preprocessing steps, a high level overview of the model architecture and key features (residual connections, spatial attention block and their purpose; downsampling and upsampling operations and progression of convolution kernels), and the outputs. 
2. The training period covers 7 years. Given the ongoing thinning and decline of sea ice cover, do you foresee a need for periodic retraining of the model? How might evolving sea ice conditions in the changing Arctic impact the model's forecasting accuracy over time?
3. Training is based on weekly datasets, yet the forecasts are daily. I presume that using weekly training data enhances the model’s generalization capability and robustness against short-term noise? However, this choice may limit the model’s ability to capture rapid, short-term sea ice dynamics occurring at daily scales. How does this choice impact forecast accuracy during periods of rapid sea ice changes? Is the reduced forecast skill during sea ice minimum periods possibly related to a temporal limitation inherent in weekly training data?
4. The evaluation spans a single year of operational forecasts. Although this period enables an analysis of seasonal performance and highlights the reduced skill during the summer, significant year-to-year variability in sea ice conditions may affect the robustness of the conclusions drawn. How confident are you in your findings after just one seasonal cycle, and could interannual variability impact where and when the model performs well? I am mostly thinking of how you might ultimately assign an uncertainty flag to the forecast data product. 
5. The authors compare MET-AICE primarily to a single dynamical model, the Barents-2.5 km EPS. How does the performance of this dynamical model compare to other available dynamical models? 

Specific comments

Line 63: It seems sensible to use 2-m temperature and 10-m winds to drive the system and you mention in the introduction that sea ice changes on short-time scales are driven by the wind. But was there any assessment of the optimal variables to train and run the model? At the very least it would be helpful to include references to justify your use of these variables to drive sea ice variability. 

Line 65: I don’t understand how the 10 different models were developed. Are each of these models for the different lead times, i.e. a set of 10 distinct forecasts for lead times of 1 day, 2 days, 3 days, all the way up to 10 days? Could you clarify the description here? Also, why do you have these different lead times - was the aim to find an appropriate lead time? Which is the dataset released via THREDDS? Is this the daily forecast with a 10-day lead time?

Line 74-75: Coastal grid points (within 20 km of the coast) are excluded from the model performance evaluation. I didn’t notice these points being masked out or flagged in some way in the forecasts released via the THREDDS server of the Norwegian Meteorological Institute. Might it be helpful to users if there is an indication of where you have confidence in the available forecast data and where users should take care.

Line 117:  It isn't particularly clear how you used the datasets from 2021-2023 and why you only produced the validation on the data from April 2024 onwards. Would having a few extra years of validation assessment have made the results more robust?

Technical corrections

Line 22: change “predict” to “predicts”

Line 203: I think “less than” should be “fewer than” in this case

This is a very well-written and timely manuscript introducing MET-AICE v1.0, the first operational data-driven short-term sea ice prediction system for the European Arctic. I particularly appreciate that the evaluation is based on the most recent year of operational forecasts, which makes the results highly convincing and relevant for end-users. The manuscript is clear, the methodology is carefully described, and the system itself represents a significant step forward for operational sea ice forecasting. The authors convincingly show that MET-AICE provides skilful 10-day sea ice concentration forecasts at 5 km resolution, consistently outperforming both persistence and the Barents-2.5 km dynamical prediction system. Overall, I find the paper well-suited for publication after minor revisions. Below, I provide some suggestions that I believe could further strengthen the manuscript.



Comments and questions

Prediction scheme design

The use of separate models for each lead time is interesting. Could the authors elaborate on why this design was preferred over a more common autoregressive approach? It would strengthen the manuscript to show why this scheme provides better sea ice concentration forecasts than autoregression. Additionally, once the 10 forecasts are concatenated, do they retain physical consistency from one time step to the next?

Forcing model evolution

At line 120, the manuscript notes the distribution shift in ECMWF atmospheric forecasts from one cycle to the next. To what extent do the authors envisage the need for retraining MET-AICE in the coming years as ECMWF forecasts continue to evolve? Have they evaluated model skill across past ECMWF cycles to quantify the impact?

Verification metric robustness

How robust is the ice edge distance error to differences in smoothing across products? I would expect a sharper forecast to yield a longer ice edge, due to more small-scale information, compared to a smoother ML forecast. Clarification here would be valuable.

Sensitivity to different forcings

Have the authors tested MET-AICE under different atmospheric forcings than those used in training (e.g., AROME-Arctic)? Would the model’s forecast skill improve or deteriorate in such cases, as seen in Barents-2.5?

Relatedly, it would be instructive to assess MET-AICE in controlled “idealised” experiments not present in the training data (e.g., constant zonal winds, uniformly melting temperatures across the domain). Do the forecasts behave in line with expected sea ice physics?

Along these lines, has the system been tested using ECMWF ensemble members in addition to the control forecast? Does MET-AICE produce a reasonable spread in SIC forecasts under such perturbations?

Missing Literature

I recommend citing also these papers: https://doi.org/10.3389/fmars.2023.1260047 and https://doi.org/10.1029/2024JH000433. Even though the latter does not pertain to the Arctic, it’s still a good example of a data-driven prediction system targeting sea ice.