The study reflects the lack of adequate reflection on previous work and literature in this area. However, this should be the basis and first step of developments and publications. This lack leads to misleading interpretations of the results and introduction of concepts as novel, which are in contrast well known. I cannot recommend this manuscript for publication in the current form. Please see the detailed comments for further information. I can not recommend the manuscript for publication, see comments below.  However, the evaluation results might motivate the publication of the manuscript after a proper major revision.

Major comments/concerns:

L105 and others:

The author uses “~16M combinations of simulated GHI at the earth’s surface”. However, several concepts exist with smarter solutions, e.g. with the hybrid eigenvector concept the amount of needed RTM calculations can be reduced to several hundred, also the good old NREL etc. approximations works well, see https://doi.org/10.1016/j.rse.2009.01.012 or https://doi.org/10.3390/rs4030622 and references therein for further details. A respective discussion or proof of the benefits of the approach presented in the manuscript is missing. Thus, it is not clear why this  algorithm is needed and what  the benefits compared to other well established methods are.

L 125 “….Wdata (Bhartia, 2012) based climatology) and surface albedo (GOME-2 database (Tilstra et al., 2017, 2021))

Several important aspects are not discussed which affects the accuracy of the product. Is the used SAL consistent with that used for COD ? Further, is the aerosol information used for the COD retrieval identical to those used for SIS  ? Are the BRDF corrections for SAL and COT identical ?  If not, what does that mean for the consistency of the product. Please note, in particular inconsistent SAL data can lead to a significant bias in SIS, this could be the reason for your bias and not the “sun obscuration”.

L137, Eq.1  

The so called Cloud Modification Factor is the good old clear sky index or cloud coverage index (Cano et al), used in several EU projects and SAFs long ago, ranging from SODA, Satellight to Heliosat-3 and CM SAF. Respective references should be given. Further, by introducing this factor to correct bias resulting from COD they proof that the direct path (see https://doi.org/10.5194/amt-15-1537-2022 and references therein) is more favorable.  It is not clear why the authors have chosen the indirect path.

L189:

Optical flow method “We apply Farnebäck”. Is Farneback still state of the art for SSI nowcasting ? In Urbich et al 2018 (https://doi.org/10.3390/rs10060955) evidence is given that TV-L1 outperforms Farnebäck. This finding is supported by other studies and the maths. TV-L1 is more robust concerning changes of the intensities.  It is part of OpenCV and thus free software as well. Further, it is not enough to compare your nowcasting method to persistence. It should be compared with other methods as well, including state of the art NWP (see respective publications of the IEA framework). Within this scope the effect of changing intensities on the quality of the forecasts should be discussed.  One of the first works in the area of solar surface irradiance (SSI) nowcasting can be traced back to Lorenz et al. (University of Oldenburg, now ISE) and others. These works should be cited and discussed as well.

L 315 “The interesting part is that the same case stands for the whole range of measured GHI, indicating that it is a general limitation of satellite that it cannot take into account clouds.” as well as L 334.  “Cases with partial cloudiness and the sun obscured as seen from the ground sensor (almost total attenuation of direct irradiance) will be associated with low measured irradiance that cannot be captured by the model. This is the main reason of the overall model overestimation."

Misleading discussion and interpretation. The satellite can of course take into account clouds that obscure the sun. Else, all values would be clear sky values, or ?  Let us assume a partly cloudy pixel with 50 % clear sky and 50 % cloudy sky, leading to an average cloudiness of 0.5. If only the sun obscured regions (100% cloud sky) are investigated you surely will find a bias, namely,  an overestimation of  SSI by the SAT retrieval, because the area average seen by the satellite is partly cloudy (50%).  But statistically, there are also situation where the ground based station sees the sun (100% clear sky), but the satellite is partly cloudy sky (50%). On average there is an “error” cancelation of these effects.  No figure or statistics are shown for situations where ground measurements see the sun, but the pixels are partly cloudy. Hence, there is no proof that the overall bias results from the “sun obscured” effect.  In several studies bias values are reported for algorithms without sun obscuration correction, which are not significant or depending on the method positive or negative (see e.g. validation reports and publications of CM SAF, e.g. Uccarra et al, https://doi.org/10.1016/j.rse.2017.07.013). Thus, the cancellation of the “errors” induced by different viewing geometries seems to work well and they are several other reasons for the bias. Thus, your conclusion seems a bit hasty and misleading.  You should check SAL, it is likely a source for your bias.

In addition, there already exists a lot of publications dealing with broken clouds, 3-D cloud effects or the uncertainties arising from the comparison between ground based and satellite based SSI. Please read them, discuss and cite them and clarify what your work adds to existing woks. In my opinion currently not much, beside misleading conclusions. Of course, for slant geometries the cloudiness is overestimated, but that is another story, which is not taken into account in your study. You will find respective articles, e.g. in the CM SAF publication list.

The aerosol study is well done, but also this part lacks a bit on discussion and citations concerning former works

Minor comments:

L54  ,”..considered as big data”.  Please delete, it is not really big data compared to other fields….

L 47 “The availability of solar resources is primarily affected by clouds and aerosols (e.g., Fountoulakis et al., 2021; Papachristopoulou et al., 2022).”

This is misleading.  In areas with low aerosol variability water vapor is much more important than AOD (as a climatology value works well there). Please add H20 as important variable.

L95 “SENSE2 is an operational system that produces fast estimates of GHI in real time every 15min, for a wide area including Europe and Middle East-North Africa (MENA)”

Please mention how the user can get these data.

L113: The aerosol model of Shettle is used, but no discussion of the limitation induced by the assumption of spherical aerosols is given.

L 140: Use of  NWC SAF products: I did not understand the sense of this approach, why do you need NWC-SAF ?   

L 145: Typical values for the effective radius (Reff = 10 μm) and the liquid water path (LWP = 1 g/m3 145 ) were used, given the unavailability of those data and their small impact on GHI"

This phrase is quite misleading. First of all there are algorithms available to derive Reff and LWP, further I would not say that the impact is small, in  particular when considering ice clouds.

L 195: "Smart persistence": I find the term irritating, please delete it.  Please clarify that this kind of persistence is typically used for SIS nowcast comparisons. Add some discussion and references of former works here as well.

The manuscript describes improvements to two high-spatial resolution models used for the prediction of the surface global horizontal irradiance (GHI) over the area of ​​Europe and Middle East-North Africa. The two models, in particular, are:

• SENSE2, a nowcasting system based on look-up-tables (LUTs) calculated using libRadtran radiative transfer model that uses as input the cloud optical thickness (COT) obtained from Meteosat Second Generation (MSG) satellite and aerosol optical depth (AOD) predicted by the Copernicus Atmospheric Monitoring Service (CAMS);
• NextSENSE2 is a short-term forecast (up to 3 hours ahead) system using the GHI of SENSE2 and the CMV technique for forecasting the satellite-derived COT.

The two model performances are validated against ground-based measurements of GHI carried out in sites belonging to the Baseline Surface Radiation Network (BSRN) in the area covered by the models and two additional sites in Greece. Measurements refer to 2017.

The analysis is mainly aimed at investigating the role of aerosols and clouds, atmospheric factors with large spatial and temporal variability, on the estimated GHI.

The prediction of short- and very-short-term GHI is one of the fundamental issues related to the efficiency of renewable energy-based systems, and the study described in this manuscript is in principle useful in supporting the development and optimization of these systems.

I think the manuscript should undergo major revisions before publication, addressing the issues highlighted as major and minor comments.

Major comments

Any references to published paper describing similar GHI prediction models or investigating the role of clouds and/or aerosols on GHI nowcasting/forecasts are missing. Thus the reader in not able to understand the goodness of the performance of the models presented.

A description of similar models should be presented in the introduction, as well as in the “summary and conclusions” paragraph the results of this work should be compared with those of similar studies conducted in the same study area or in different regions.

The performance of the nowcasts in paragraph 3.1.1 is not well supported. The sentence “This overestimation is attributed to the underestimation of cloud related information from satellite (MSG COT), when we compare point measurements with a pixel in satellite images corresponding to a wide area of almost 5 km x 5 km” needs to be argumented because no evidence of COT underestimation is supported here.

Moreover, the authors attribute the model's overestimation of BSRN measures for low GHI values ​​to stations with more cloud cover, particularly those at high latitudes, such as Lerwick. However, evidence of the cloudiness in the various sites is not provided and the results are not presented for a single station. In my opinion a GHI scatterplot similar to that of Fig 4a for individual stations could be added as supplementary material.

The discussion of paragraph 3.1.2 on the aerosol effects on cloud-free GHI should be completed with the appropriate references addressing the CAMS and AERONET AOD comparisons.

Minor comments

Line 30:  use “significantly improved” instead of “improved a lot”.

Line 51: add a sentence on the large temporal and spatial variability of clouds and aerosols.

Line 70: change “form” with “from”.

Lines 71-73: is there a reference to cite for this sentence “The validation of this method showed a good agreement on daily and monthly levels; however, various sources of uncertainties have been identified, concerning mainly the use of NN especially during high irradiance atmospheric conditions, the COT, and the structure/density of atmospheric parameters in the LUTs”?

Lines 81-83: the meaning of the sentence “However, this first evaluation was based on the satellite-derived COT, so the aim of the current study is to compare the irradiance forecasts with ground-based measurements.” is not clear.

Line 96: is there a web link to reach the model and see the GHI estimates? Similarly for NextSENSE2. In case it is useful to add it.

Line 109 and line 112: put a space before “nm”.

Line 129: briefly explain how to correct the surface GHI for sites at higher altitudes than sea level.

Line 145: COT, R_eff, and LWP a strictly related. The simplest way to see the relation is the formula.

LWP=C*r*COT*R_eff, where C depends on the assumption of the R_eff vertical distribution within the cloud, see e.g. Wood and Hartmann, J. Climate, 19, 1748–1764, https://doi.org/10.1175/JCLI3702.1.

So if COT is allowed to change in the RTM model simulations with R_eff kept fixed, LWP can not remain fixed to 1 g/m³.

Line 147: the cloud cover fraction is one of the RTM input variable. How is it treated in the simulation of the LUTs?

Line 195: Are there any approaches different from the persistence one to account for modifications in the cloud optical and physical properties?

Line 204: some little information and reference for the two non-BSRN sites of Athens and Thessaloniki may be added.

Line 208: How is the clear-sky GHI derived for non-BSRN sites? Do authors know how well the Ieichen-Perez clear sky model performs? Did they estimate the deviations compared to GHI measurements in cloud-free conditions?

Equation 6: rRMSE_CMV and rRMSE_pers. are not introduced.

Line 242: “due to the limitations in the field of view of the satellite”. Explain.

Line 305: “CMF>0.9” is “CMF≥0.9”.

Line 308: use “0.4<CMF<0.9” “instead of “CMF <0.9 and >0.4”. This is valid for the rest of the manuscript.

Line 309: change “the lowest values of measured GHI are found (<250 W/m²)” with “the largest occurrence of small measured GHI values (<250 W/m²) are found”.

Lines 310-311: again, how do authors support the MSG COT underestimation? If it effect is more evident for high latitude sites, this should be shown.

Line 327: report the MBE.

Line 335: until now the authors have not mentioned the 3D effects of clouds and the fact that these cannot be reproduced with 1D models, especially in conditions of partial cloud cover. They should mention this as a limitation and cite the appropriate references.

Figure 7a: the figure could be larger and the text inside the graph is hard to read.

Line 377: the authors mean that the MBE and RMSE are improved after correction, as it is obvious.

Line 383: I would have expected a greater increase in cases with GHI differences within ±50 W/m² after correction.

I suggest a general review of the English language.