This study introduces a short-term nowcasting model that combines techniques such as machine learning and data assimilation in a novel fashion to help in predicting the direct normal irradiance. Validation of the models and methods is thorough and the authors take the time to explain the interpretation of their results. The addition of an “undecided” class to the training step is clever, especially when tied to their loss function which works to sidestep uncertainties in classification. Data assimilation from two separate imagers is used appropriately and adds an extra layer of context for the initial state. Overall, the paper advances the state-of-the-art of nowcasting by combining several innovative methods and could serve as a baseline for future research in the radiation or energy communities using such techniques.

Specific Comments:

1. The caption for Fig. 2 needs to be more descriptive. As well, Fig. 2  is not adequately described in the main text when it is first referenced although lines 251-253 do add more context. I would strongly suggest providing the reader with that context for the figure to start with by adding more information in the caption.
2. In Sect. 3.2, it would be helpful if a histogram of the retrieved cloud base height is added. This would allow a reader to quantify the performance of the derived base height for the entire set of scenes without limiting samples as has been done for Fig. 3.
3. Starting from line 340, the authors use a value within parentheses when describing the irradiance. It is unclear as to what these values are referring to, particularly as there is a preceding value before the parentheses as well. For instance, lines 343 - 344 say “Typical improvement over persistence for these longer lead times is thereby on the order of 50Wm⁻² (50Wm⁻²) and more” but both values being the same creates confusion. I would recommend introducing the parameter within the parentheses first or explaining it at the top of the paragraph.
4. There are a number of grammatical errors overall that will need to be corrected before publication. For instance, in lines 20-21, the phrasing should be “Since direct irradiance can be blocked completely by clouds within seconds to minutes, knowledge of future direct irradiances is especially important for solar energy applications.”. The incorrect use of adverbs and articles in many places interrupts the flow and might particularly detract a reader from the point of the sentence or a paragraph which is why I am adding this issue as a major comment. 
5. In the appendix, it is mentioned that the ResNet encoder of the CNN uses pre-trained weights from ImageNet. This seems like an unnecessary step as the ImageNet classes are oriented at natural object detection (and not clouds) and transfer learning from a pre-trained ResNet would not necessarily reduce the convergence time on a task such as cloud detection. Could the authors clarify why a pre-trained model is better as opposed to training from scratch for this application?

Technical Corrections:

1. The LaTeX equations have not rendered correctly in the preprint. For instance, line 207 has a question mark instead of equation numbers. This needs to be corrected.
2. The cloud optical depth threshold in line 273 should be reversed to say τ>τ_thresh is classified as a cloudy region.
3. In line 388, there is mention of mean absolute error. Since this metric is not presented in the main text or appendix or supplement, I would recommend removing this sentence as it is unnecessary. The RMSE and MBE already provide sufficient quantification.

The presented study is a valuable addition to the increasingly important topic of very short-term solar irradiance nowcasts. The methods used in the presented model are a mixture of existing and new approaches. 

Here, the data assimilation approach derived from numerical weather prediction for an optimal initial state is very promising. Furthermore, the validation strategy using synthetic ASI images derived from LES simulations should be positively highlighted. This enables the comprehensive validation capabilities as mentioned by the authors.

The manuscript is coherently structured and explains the applied methods sufficiently

I can agree with the comments of the previous referee.

General comments to improve quality:

• When reading the manuscript I was sometimes confused about synthetic and real data. E.g. readers might associate "DNI measurements" with instrument (pyrheliometer) measurements.  It was also not clear to me if ASI images for training the CNN model were real ASI images. Please take care of a clear distinction when real and synthetic data has been used 
• The decision of choosing DNI vs. GHI or DHI can be explained in more detail. The estimation of spatial distributed diffuse horizontal irradiance (DHI) from ASI is a generally rarely addressed issue in ASI based nowcasting and worth to be mentioned. DHI is also needed for transposition modelling needed for solar energy applications. I can also imagine that the used setup with MYSTIC RT modelling and LES is capable in addressing DHI estimations. 
• Chapter 2.2.2: The quality of the used data assimilation can be better explained and is probably of interest for the reader. A graphical example of model, observation and analysis states might be a solution for this.   
• Persistence comparison: in general, by definition, the quality of persistence decreases with cloud variability. The syntethic data set with increasing cloud cover is therefore an ideal basis for an investigation of forecast skill of the model vs. persistence. Therefore, a comparison of forecast skill with respect to cloud fraction would improve the manuscript.

Technical comments:

General: The figures are a quite small and therefore difficult to read. Possibly they can be made larger for the final version.

Line 3: You might add that the CNN was trained on "real ASI images" to avoid confusions. 

Line 27: It should be Eye2Sky instead of eye2sky

Line 195: Please check k,l it might be mixed with l,p as defined in the equation above

Line 330: The definition of cloudmask variation and continuous cloudmask variation might be introduced here. It takes a while to understand the difference

Line 338: grouped instead of groupped

Line 347: RT definition not introduced