The efforts of co-authors in addressing the reviewers’ comments are appreciated. There are however a couple of concerns that are not appropriately answered in my opinion.

Both referees where concerned by the number of scenarios required to train the ERSM and there remain a couple of unsolved issues. Please provide a more detailed explanation for the reason why the number of required scenarios would be of a power of four of the number of variables in the present case. With 10 to 15 variables, that would be a maximum ensemble size of tens of thousands, not hundreds as stated in the manuscript (P4 L13). While reducing that size to 600 is a very welcomed improvement, it remains a substantial number and it cannot yet be claimed that the proposed technique is “highly economical”. To address this point, and while none of the reviewer requested a more detailed quantification of the convergence rate, the authors may consider including a figure illustrating the evolution of the mean normalized error when increasing the number of simulations.

A short summary of the reason why « Xing (2011) indicated that the nonlinearity in atmospheric responses could not be captured in metropolitan regions unless fourth or higher order equations were used » (P3, L29-31) should also be included since that reference is in the grey literature, and in Chinese.

Both reviewers requested more details on the rationale and implication of neglecting interactions of transported and returning precursors, this hypothesis is indeed now better stated in the revised manuscript (P8L9-10 and P9L11-12). However, the possible implication of such an hypothesis is not given in the revised manuscript. Can you think of a possible quantification that would support neglecting this process?