Govardhan et al. introduce DSS1.0, an operational air quality forecasting and emission scenario framework with high resolution WRF-Chem simulations at its core along with other enhancements such as chemical data assimilation from ground-based stations and satellites to improve initial conditions and tagged tracers for different sectors and regions with different emission levels. Their forecast system has the novelty of providing real-time source apportionment of PM2.5 along with the forecasts as well as sufficient tagged tracers from different sectors and regions to build bespoke short-term episode-specific mitigation policy. This is commendable work and should be very attractive for policymakers.

The paper is well-written with clearly stated objectives. There are also some notable details which show the scientific robustness of the work, for example, the details of the chemical data assimilation as applied to the tagged tracers, and the “feedback” module in the forecasting system which adjusts the new initial conditions after an actual policy is implemented in the real world.

However, I’d like to point out some weaknesses which need to be addressed before the work is published:

1. The authors explained the underestimation of PM2.5 concentrations in the first half of November due to the lack of firecracker emissions in the model during the Diwali festival, which is understandable, but they haven’t explained the dramatic overestimation around the third week of November (even if the policy intervention was not included in the model). Is this due to wrongly generated timing of fire emissions or a poor meteorological forecast during that week? This needs to be discussed in sufficient detail.
2. The forecasts completely miss three large observed peaks in winter (Fig 5c). This surely cannot be due to poor representation of biomass burning in the model because there’s hardly any agricultural burning going on during this period in the region. The authors have touched upon the issue of lack of a dynamic emission inventory, but can these peaks be captured just by applying a temporal cycle to the existing emissions, or are some key emission sources/processes missing in the model? The authors need to discuss this in more detail. I’m thinking of open-waste burning, brick kilns, gas-to-particle conversion etc. but a bit more local knowledge needs to be added here.
3. The defence that the source apportionment results shall hold true on a percentage basis even when the model underestimates the episodes is only partially true. This is because, the relative contributions of local and near-regional sources might disproportionately increase during the the peaks which will not be reflected in the source apportionment results if those peaks are not captured. This is of utmost importance because these are exactly the periods when a policy implementation might be wanted. Therefore, the authors need to acknowledge and discuss this weakness in the forecast system and propose potential solutions.
4. The aerosol module used in WRF-Chem doesn’t represent secondary organic and inorganic aerosol production. This is understandable and the authors have defended their choice well given the computational constraints of running a nested high-resolution model along with several tracers especially when PM2.5 isn’t directly simulated, and they had to tag various PM-components for each sector and region. However, lack of gas-to-particle conversion in the model will have an impact on the contributions in the scenarios. For example, in the real world, traffic emission reductions might lead to a significant reduction in nitrate aerosols, but this won’t happen in the model. The same goes for energy/industry emissions and sulfate aerosols. In that sense, the air quality improvements from the scenarios in the DSS might be underestimated – this needs to be clearly discussed.
5. For Figures 7 and 8, which of the forecast has been used: 1-day, 2-day, 3-day, 4-day or 5-day forecast? This needs to be stated in the figures.
6. When suggesting the policy recommendation based on scenarios, I suppose the 4-day or 5-day forecast will be more practical than the 1-day forecast as it will allow some time for decision-making. However, the 4-day or 5-day forecasts will have a 20% or 40% reduction throughout those 4 or 5 days based on the tagged tracers, while the policy might be implemented at a later start date – this may lead to discrepancies in the outcomes. Therefore, the authors need to make this point clear.
7. A proper evaluation of this system should include not just the forecast performance against observations but also the accuracy of scenarios. Therefore, the policy intervention that happened post-Diwali should ideally be evaluated against the closest possible scenario based on the tracers. Does the implementation of the closest possible emission scenario resembling the post-Diwali policy intervention successfully reproduce the drop in PM2.5 during that period? This would be a real litmus-test of the system, and even if it doesn’t reproduce the drop accurately, we will at least learn about the discrepancies and get a step closer to making the modelling system reflect the real-world conditions. Such an evaluation will inform emission inventory modifications or new process representations. Therefore, it would be very valuable for the community if the authors can perform this evaluation.

Once again, I commend the authors for this crucial and significant work, and I have no hesitation in recommending it for publication in GMD once these issued are addressed.

General comments

The manuscript “Decision Support System version 1.0 (DSS v1.0) for air quality management in Delhi, India” by Govardhan et al. introduces a model system for short-term air quality forecast and emission reduction strategies in Delhi during the post-monsoon and winter seasons of 2021-2022. The authors use the WRF-Chem model with specific emission inventories to forecast the regional PM_2.5 concentrations and Air Quality Index (AQI) over Delhi in five days, and they add some tagged-tracers to quantify the contributions of emissions from different sources over the local and surrounding regions of Delhi. The authors also design two scenarios of emission reductions to evaluate the impacts of sectorial anthropogenic emissions on the PM_2.5 levels in Delhi.

In general, the manuscript is well organized and written, which fits the scope of the Geoscientific Model Development. The proposed model system in this study is promising to warn the short-term air pollution events and it’s useful for local policymakers to manage the air quality in Delhi. However, it would be better if the authors can describe and discuss more details on how to estimate the biomass burning emission inventory, as well as its uncertainty and impact on forecasting PM_2.5. It would be more convincing if the authors can provide some observations of PM_2.5 compositions to evaluate the model simulations, as they use a relatively simple aerosol scheme lacking some secondary aerosol species. The authors might uniform and enlarge the labels and legends in figures for a better reading experience. The reviewer recommends publication after major revisions. Please see the specific comments and technical corrections listed below.

Specific comments

P3, Line 130: Which version of WRF-Chem do you use? Please add the citations for the WRF-Chem model.

P4-5, Line 150, 169, 193: Please add the citations for the IITM GFS, EDGARv4.3 inventory, and the MODIS active fire count data.

P4, Line 179: The authors use the anthropogenic emission inventory from TERI for the year 2016. Is there any increasing or decreasing trends of anthropogenic emissions from 2016 to 2022?

P5, Line 192: I would suggest the authors to give more information on the forecasted fire emissions. I’m wondering if this method can capture the day-to-day variability of fire emissions in a short forecast period based on the climatological fire emissions.

P6, Line 260: I’m wondering if lack of nitrate and ammonia aerosols in this model would cause the biases in total PM_2.5 concentrations. It seems that Figure 2 is based on the WRF-Chem simulation. I would suggest the authors to show some observations of speciated PM_2.5 concentrations or cite some previous studies.

P11-12, Section 3.1: The authors explain the underestimation of PM_2.5 concentrations in the first week of November is due to the large uncertainty of fire emissions. As I mentioned above, I would suggest the authors to discuss more about the method on predicting the daily fire emissions, because the daily variation of fire emissions may vary with the weather conditions and some human activities, which may introduce the large biases during the burning seasons. Could the authors show the estimated fire emissions during this period? I’m also wondering if the overestimation of PM_2.5 concentrations in the following weeks is caused by the inaccurate anthropogenic emissions or the fire emissions.

P16, Table 4: Why does the POD for the “Poor” AQI category decrease by 30% to 40% in the Day 4 and Day 5? But the Accuracy doesn’t change a lot. I would suggest the authors to give some thresholds for these statistical parameters indicating the confidence level and reliability of the DSS system, which may be more helpful for the policymakers. Maybe the authors can add the reference curves in Figure 6.

P20, Section 3.3: The two reduction scenarios show a very good linearly relationship as the authors include very little secondary PM_2.5 species in the DSS system. I can understand that the authors use the GOCART aerosol scheme for computational efficiency. But it would be better to implement some previous studies on the contributions of secondary PM_2.5 to total PM_2.5 in Delhi.

Technical corrections

P1, Line 36: Please spell out the acronyms “NCR” when it first appears.

P6, Line 265: Change the “SO₄^-2” to “SO₄^2-” in equation (2).

P7, Line 284 and Figure 2: Change the “SO₄^--” to “SO₄^2-”.

P20, Line 682: Change “suggests” to “suggest”.

P21, Figure 8: Please make the label of x-axis clearer. The “Day-month” is a little bit confusing. I would suggest use the date for x-axis.