The paper is well written. The authors applied an advanced deep learning architecture and conducted a detailed and interesting sensitivity study on the predictability of ozone using the U-net+LSTM. Analyses on the forecast lead time, the added value of upstream history, and the concept drift could be useful for future studies. The inclusion of reference models enhanced the validity of the results. I recommend the publication of the paper with minor revisions.

General comments:

• At a resolution of ~12 km, could you comment on the potential impact of sampling errors, when your model is applied to real observations in the future? How many sites do you have in each grid box?
• Could you discuss the ozone chemical regime over Germany and the impact of VOC on ozone predictability?

Specific points:

L161: Inconsistent font of “DataHandler”

L278-280: Is it possible that the training/test performances are different because your NN model is slightly overfitted towards regions with more pseudo-stations?

Figure 9: Could you comment on the degraded performance over the coastal regions? Do these coastal sites have something to do with the results of skill scores in Figure 10? For example, the north wind driven by the sea breeze circulation brings less useful information.

Figure 11-14: How would the learned non-linearities compare with the WRF-Chem CTM? Could you comment on that?

L330: I think your NN3s model slightly outperforms OLS at most of the pseudo-sites?

L340: I see measurement uncertainties are not utilized for now. Can you comment on potential utilization of measurement uncertainties in the future (e.g., via Bayesian NN)?

This paper presents an interesting work on the use of machine learning to forecast tropospheric ozone. The current version answers most of the previous review remarks. However, for the sake of completeness, I would raise two questions that could be addressed in the text as "discussion" even though some of them could lead to more experiments in future works. 

Indeed, the first remark is related to the forecast model chosen by the authors, that creates time-serie forecasts for each individual air-quality station. As discussed in the text, the experiments show that better results are obtained when using data from surrounding stations rather than a single one, which seems logical as this "nighborhood" strategy helps capturing the advection ozone. For this reason, it would be interesting to develop a bit more the comparison with models that use spatial data, as for example https://doi.org/10.1007/s10994-020-05944-x that uses 2D matrices of the entire area in the form of images and video frame prediction algorithm. I believe that WRF-Chem outputs can easily be represented as 2D matrices.

The second question concerning the datasets. In my opinion, only a single year seems too limited to catch the dynamic of the atmosphere. As training, validation and test subsets cover different months of the year, seasonal variations are certainly present and cannot be correctly expressed. A more clever strategy would be to use datasets from several years (even if only in a reduced number of months per year), and validate with the same period in another year.

As stated before, answering these remarks is not mandatory but would greatly contribute to the paper completeness, even if just in the form of a discussion.