This manuscript presents a machine learning technique for reconstructing ionospheric profiles from ionograms and geophysical parameters, using incoherent scatter radar data as the truth data for training. The explanations in the paper are very clear, and the final results are convincing. I have a few minor comments for the authors to address.

• Use of ARTIST 4.5: The manuscript on line 120 says that the current processing for the Tromso ionosonde uses ARTIST 5.0 and ARTIST 4.5 has been discontinued. Why then do all the comparisons in this paper compare to ARTIST 4.5 not 5.0? Furthermore, this paper shows that ARTIST 4.5 performs very badly, to the point that it has a negative R^2 coefficient. What is making ARTIST 4.5 so bad in these comparisons? Line 293 says missing traces are being filled in with 1e8 m^-3. Are those low placeholder numbers artificially wrecking the R^2 for ARTIST 4.5? I would expect ARTIST 4.5 to at least get the peak density correct since it should extract foF2 correctly.
• Figure 10 panel g shows that at certain times, none of the techniques can reproduce the measured profiles and all R^2 drop to zero or lower. What is happening during these intervals where none of the techniques work?
• Line 81 implies that only dates with a full 24 hours of coverage were used. Why is 24-hour coverage required if the model is being trained on individual profiles? This methodology appears to discard a lot of potentially usable training data.
• Why was 2019-2022 selected, and why was a larger portion of the available data not used for training? I am concerned about the biases in the training data, both including biases towards solar minimum and biases towards winter months.

The work addresses an important challenge in ionospheric research by introducing a multimodal deep-learning approach (KIAN-Net) for estimating electron density profiles from ionogram images and geophysical parameters. The manuscript is well written, easy to follow and reads well. The methodology is sound, the analysis is thorough, and the comparison with established models (ARTIST 4.5, E-CHAIM) is both relevant and convincing. The figures are clear, and the results are significant for both operational and research contexts. The study’s novelty lies in combining ionogram and geophysical data in a unified neural network framework and benchmarking its performance under auroral and disturbed conditions, where traditional scaling techniques struggle.

I have only a few suggestions and clarifications that could strengthen the paper.

In the discussion, the authors note that ionograms dominate the predictions compared to geophysical parameters. This is an important point that could be supported quantitatively. I recommend performing (or at least discussing) a feature importance analysis, e.g., SHAP values or permutation importance, to quantify the contribution of each geophysical parameter to the model output. Such an analysis would reinforce the justification for including the geophysical parameters and could reveal which indices are most relevant to the model.

Line 40: Among the referenced scaling techniques, the author should also include Autoscala (Pezzopane & Scotto, 2007).

Appendix A: In the reviewer’s opinion, the outlier detection and correction are not completely clear in their methodology.

• I think that PCA should be explained further for readers who are not familiar with typical preprocessing steps in machine learning. In the reviewer’s opinion, a brief description of PCA would thus be helpful.
• The authors say that outliers are values falling outside the IQR range. What does value refer to? Is it an electron density value at a certain height? Is it the whole profile? Over which values were the IQR and median calculated?
• Outliers are removed by applying a median filter with a kernel size of 5. What does this mean? Is the median filter the same as that which was applied to smooth the ISR images shown in Figure 2b? Do you apply the median filter to a time series of electron density at a fixed height (rows of the image shown in Figure 2a)?
• According to the methodology, outliers are removed by performing median filtering with a kernel size of 5. As a median (not mean) filter, the reviewer expects the filtered profile to match one of the values of the previous or following 2 profiles shown in Figure A1. Are the profiles shown in blue\green already filtered with PCA? Is the filtered one also gone through PCA?

Pezzopane, M., and C. Scotto (2007), Automatic scaling of critical frequency foF2 and MUF(3000)F2: A comparison between Autoscala and ARTIST 4.5 on Rome data, Radio Sci., 42, RS4003, doi:10.1029/2006RS003581.AQ