It can be observed that the reviewers are divided on the merits of this work. In addition to the public comments to the authors, there have been a number of private messages to the editor that I have taken into consideration.

Probably the most problematic aspect of this amongst the more negative reviewers is the use of fixed scaling factors based on concentration, which seems arbitrary and could be expected to work better for some pollutants than others, specifically ones dominated by primary emissions (e.g. NOx) rather than secondary (e.g. PM2.5). I would also observe that there seems to be an underpinning expectation that PM2.5 should be expected to follow NOx and SOx emissions, yet it has long been established that the relationship between PM2.5 and those precursors is much more complex than that, so a deviation here after this treatment should not be in any way surprising.

While the authors do admit to such limitations in the paper, this throws into question the quantitative validity of the results. The qualitative impact of COVID restrictions on ambient pollutant concentrations is not in doubt, however the purpose of modelling papers such as this is to try to attach more quantitative estimates of the changes in the emissions, so if that aspect is questionable then this speaks to the scientific significance of the work.

Taking all of the reviewers' comments in balance, along with private communications, I am not ready to reject the paper as fundamentally flawed, however given how unconventional and evidently contentious the approach is, I have decided that the authors must perform more work to critically assess its validity and significance of the conclusions arising from it. As such, I will reconsider this after another round 'major' revisions.

What I would suggest is that the authors pay more attention to the limitations as well as the benefits of their approach and make sure that the conclusions arising are presented in a balanced manner accordingly. As well as suggestions for revisions made in previous reviews, there were also potentially helpful suggestions made in private communications. These included the splitting of the dataset into 'training' and 'evaluation' datasets, or perhaps treating the different pollutants as separately when reaching the conclusions.

It seems that the main point of contention - the use of the beta factor - has now been satisfactorily addressed. One reviewer has a few comments, although I do not regard these as 'major'. I would invite the authors to respond to these.