The submitted paper, titled "How to trace the origins of short-lived atmospheric species in the Arctic", investigates the origins of particles and trace gases in rapidly changing polar climates, with a focus on aerosol-cloud interactions. The authors highlight the limitations of current backtrajectory models like FLEXPART in identifying emission sources accurately, emphasizing the need for improvement due to the impact of aerosols on polar clouds and climate modeling.

To address this, the study combines backtrajectories from FLEXPART with tracer simulations from WRF-Chem, enabling a precise evaluation of source detection methods. They present a new approach based on backtrajectory analysis, to improve source identification accuracy through parameter sensitivity studies and validations using Arctic aerosol campaign data. Their results demonstrate the flaws of traditional backtrajectory analysis and the skills of the revised method in correctly identifying sources of methane sulfonic acid and black carbon.

The methodology presented in this paper appears robust and well-developed, addressing key challenges in tracing the origins of short-lived atmospheric species in polar regions. The combination of WRF-Chem tracer simulations and FLEXPART backtrajectory analysis represents a significant step forward in improving source identification accuracy. The results are convincingly validated with observational data, making this study a valuable contribution to the field of atmospheric sciences.

I recommend this paper for publication, subject to the authors addressing the minor comments outlined below

1) The choice of a 50 km × 50 km grid resolution for FPES calculations might limit the method's ability to resolve emissions from localized or highly dynamic sources such as ship traffic. Given the transient and narrow spatial footprint of such sources, the averaging approach inherent in the method could dilute the contribution of mobile emissions and introduce overlap with nearby stationary sources. Have you tested the sensitivity of your method with a 25x25km or smaller grid?

2) The optimized cutting threshold on FPES is set to 2%, which the authors appropriately note in the discussion cannot be generalized to other regions or surface sources. To enhance the potential for generalizing these results, I suggest also to translate the FPES values into a quantifiable number of trajectories contributing to a given FPES value. This would provide a more universally interpretable metric for future applications. When stating that the purpose of filtering the FPES is to remove isolated backtrajectories, could the authors clarify what constitutes an 'isolated trajectory'? Does ir refer to a single trajectory, or a minimal number of trajectories in a given grid cell?

3) Could the authors elaborate on how their method would perform in isolating the impact of ship traffic, particularly in areas where shipping lanes are adjacent to other emission sources, such as coastal or industrial regions? Would finer grid resolutions or additional filtering parameters improve the reliability of source detection for such cases?"

Summary

The manuscript 'How to trace the origins of short-lived atmospheric species in the Arctic' by Anderson Da Silva and co-workers analyses the robustness of a previously used, qualitative source attribution (localisation) method by applying it in what would be called an observation system simulation experiment (OSSE) in the inverse modelling community. They can show that the method suffers from several limitation, which they then try to remove by applying additional filtering (baseline subtraction, removal of areas with low residence time). Similar techniques have been used in previous studies, which is not reflected in the current manuscript and gives the impression that the 'improved' method is wholly novel. Nevertheless, one novelty of the 'improved' method is the combination of high concentration footprints with low concentration footprints. Furthermore, the application of the 'improved' method to real observations of two atmospheric compounds confirms the general suitability for the presented setup of Arctic measurement stations and the compounds treated (short-lived with exponential decay). Although, the manuscript is mostly clearly written it suffers from a lack of clarity on past and present model approaches for source attribution, both in terms of awareness of existing methods and description/application of transport simulations. Furthermore, the shortcomings of qualitative source attribution methods are well known in the greenhouse gas and air quality community and, hence, such methods have largely been replaced by quantitative inverse modelling. Although, some of these shortcomings may be addressed solely by text modifications and additions, others may require revisiting some of the analysis.

Major comments

Alternative attribution methods: The introduction focusses on source attribution methods that have been used for arctic aerosols and/or ice nucleation particles (L43f). However, similar methods have been applied to long-lived and short-lived compounds for more than three decades. In general, three types of statistical source attribution methods can be distinguished based on different kinds of Lagrangian modelling: 1) the present ratio method, which traditionally was termed 'potential source contribution function' (PSCF) or 'Ashbaugh method', 2) concentration-weighted trajectory (CWT) or 'trajectory statistics' and 3) inverse modelling. A fourth, even more fundamental method, would be the interpretation of back-trajectories or footprints (or potential emission sensitivity (PES), as used by the authors) of individual or aggregated events/samples. The three statistical methods are described/tested in the manuscript by Fang et al. (2018), which is used for motivation also in the current manuscript. Another example/review of these methods is by Brunner et al. (2012) and gives additional references to the origin of the methods. Although, both studies focus on long-lived greenhouse gases, these alternative methods need to be mentioned in the motivation. Of the abovementioned methods, 1 and 2 are purely qualitative methods that can only provide indications of potential source areas but cannot determine actual emission strengths. Both methods can be applied to output from single trajectory models and from Lagrangian particle dispersion models (LPDMs, see next comment). Method 3 (inverse modelling) is a quantitative method for source attribution (location and strength) and should be the tool of choice for greenhouse gas emission attribution, but also applied for air pollutants. The manuscript wrongly states that Fang et al. (2018) 'do not present a protocol of use capable to identify sources of the studies atmospheric species' (L51f). However, Fang et al. clearly encourage the use of inverse modelling over the two other methods. Furthermore, on L501 it is stated that 'standard ratio method is more advanced than widely used inverse modeling analysis'. That's exactly not the case, as explained above. Inverse modeling is (usually) guided by prior information and through its quantitative nature and rigorous uncertainty treatment superior to the PSCF or CWT methods. This does not mean that testing and applying the PSCF method as in this manuscript should not be attempted, but the introduction and conclusion need to reflect the state of research in source attribution methods more completely.

Single trajectory vs. Lagrangian particle dispersion models: The authors seem to treat studies carried out with (single) trajectory models and LPDMs synonymously. However, there are important differences to be considered that are not well reflected in the motivation and placement given in the present manuscript. Single trajectory methods do not reflect the dispersion of air masses at all and, hence, do not well reflect quantitative air mass transport, especially when transport in the atmospheric boundary layer is concerned. Lagrangian dispersion models describe turbulent (and convective) transport through a stochastic process on a multitude of model particles (air parcels). Applied correctly, they provide quantitative transport statistics. Consequently, source attribution methods solely based on single trajectory models suffer even more non-quantitative challenges. To make the distinction between the two types of Lagrangian models I suggest introducing them as such in the introduction and to refer to output of these models by back-trajectories and PES, respectively.

Improved ratio method: It would be interesting to see how the different modifications of the old ratio method impact the results of the improved method. Is the main improvement through the cropping of low residence time areas, the background subtraction or the additional modification through the low concentration footprint? From Fig. 3 is appears as if not too many of the selected observations for the high concentration case change from the original selection to the above-background selection. If the residence time cropping is the main cause of improvement, the conclusions drawn here may be very specific to the present setup of observing sites and source areas.

Specific comments

L8, 'commonly used back-trajectory analysis': Not sufficiently specific. If by back-trajectory, single trajectory simulations without dispersion are targeted, I would agree. Otherwise, this is too general and needs to be more specific to the kind of analysis tested in the manuscript.

L35f: There are references given for the first kind of studies mention in the sentence, but not for the second kind (correlation with chemical tracers). Please provide examples.

L42f: References given for INP studies use a mix of different methods, from single trajectories to ratio methods relying on LPDM output. These different kinds should be listed separately and references given accordingly. The publication by An et al. (2014) does not seem to contain INP at all, but focuses on CO. Why was it mentioned?

L49ff: The interpretation of Fang et al. (2018) was already mentioned in the general comment. The final sentence of the paragraph needs to be more specific again as OSSEs are routinely carried out for inverse modelling of greenhouse gases.

L65ff: The paragraph introducing the specific analysis method should be moved into the introduction, as it contains general discussion of modelling concepts and is used to motivate the present study and not to describe methodological details.

L68f: The list of Lagrangian models should be sorted by model type, single trajectory models (HYSPLIT, LAGRANTO) vs. LPDMs (FLEXPART, STILT). HYSPLIT can be run in dispersion mode as well. Another frequently used LPDM is NAME (Jones et al., 2007).

L71: Here is an example of the use of back-trajectories for the output of a LPDM. As mentioned in the general comment above, replace it with potential emission sensitivities.

L93: What is meant by 'assimilated' in this sentence? Did you mean associated? I think corresponds would work best for all tracer types listed in the sentence.

L98f: If the simulated tracers are supposed to represent aerosols, why was an exponential decay chosen over a tracer undergoing typical aerosol removal processes (settling, dry and wet deposition)? The exponential decay somewhat simplifies the behaviour of a real tracer and may represent an easier target for the source attribution than real aerosol. Please give additional motivation for this choice and discuss the limitations.

L110f: Does this mean that wet deposition was considered in the simulated tracers after all? Please clarify.

L140f: 'particle diffusion all along the edges'. Sounds odd. How an LPDM simulates atmospheric dispersion was mentioned above. If LPDMs are introduced properly in the introduction, FLEXPART can just be introduced as such and its output being PES.

L145f: Does this mean with a mass proportional to the WRF-simulated concentrations? In backward mode the mass given in a FLEXPART RELEASE is ignored and set to unity. Hence, such an approach would not, and should not, reflect the observed concentrations at the receptor at all.

L147: The approach to release/initialise model particles/trajectories in a larger area is usually not applied to LPDMs that can simulate dispersion. This is usually done when employing single-trajectory models to mimic dispersion by an ensemble of trajectories with different initial locations. It is one big benefit of LPDMs that they can be used to treat point sources properly and in backward mode this means that they can correctly represent point measurements, better than Eulerian models, where you need to interpolate to a given location.

L148: 10'000 particles per day sounds a bit low to produce robust FLEXPART simulations. Usually, release rates of ten thousand of particles per hour are used. If I interpret Irish et al. (2019) correctly, they even released 100'000 every 20 minutes. Raut et al. (2017) seem to have released 10'000 every ten minutes along their flight track.

L149f: This should not happen when a sufficiently larger number of model particles was selected. The residence time calculated by FLEXPART should not be proportional to the number of released particles, because each air parcel in FLEXPART is assigned an equal time fraction that is proportional to 1/N, N being the number of released particles. I suggest performing two test simulations for a single day and set the number of released particles to something like 240'000 and 480'000 particles per day and compare the output of such runs with each other and the previous run with 10'000 particles. If the two with the large number show less difference between each other than compared to the one with 10'000 particles, I would think there is a strong need for a note of caution to be given for the employed FLEXPART setup.

L186: Consider '… no longer the presence but the absence of sources or even the presence of sinks'.

L197ff: Is this really a fair evaluation? Since source strength is not uniform over the area but depends on wind speeds, a better comparison would be the average flux map            used in the WRF simulations.

L201: Not quite clear how D is calculated. An equation or additional description would be helpful.

L225: For the reader to understand the magnitude of the failure, it would be good to show such a failed experiment as well, similar to Fig 2.

L229ff: I don't think it is the uncertainty of the transport model that is the main reason for failure here. I would rather think that it is the relative position of source areas and receptors. Any source attribution method based on a single observing site will suffer from a so-called shadowing effect. This tends to falsely assign emissions to areas that are upwind of the true emitting area. In the present case and for the sea ice and ocean tracer, the continental areas are mostly in such a configuration. The only way to robustly overcome this problem, is a network of sites that can observe gradients across the domain or at least 'observe' the same source area under different flow directions (as demonstrated in section 3.5).

L250: Usually this is called background subtraction, which is supposed to separate slowly varying background concentrations from recently added pollution events (e.g., Ruckstuhl et al. 2012, Resovsky et al. 2021). How was this done here?

L251: Using a residence time threshold to remove 'boundary' effects has been done in many source attribution methods before. For example see the factor W_ij in Fang et al. (2018) that is based on the number of trajectories passing through a grid cell, but is nothing else than a residence time threshold when translated to LPDM output.

L265: 'leaks' sounds a bit strange in this context. I would prefer the term 'shadowing' or more general 'caveats'.

Equation 6: This is the innovative part of the method and it should be mentioned as such.

L274: What is Rl_uparrow_33 set to elsewhere? In areas where it does not take the value of Rl_33.

L278: What 'bars'? Isn't this simply the number of correct detections/attributions?

L291ff: It seems that this paragraph describes figures that are not shown in the manuscript. Consider adding to a supplement or annex.

L300f: A more suitable example in this context would be the networks for greenhouse gas observations that are used to attribute global and regional emissions/sinks (e.g., ICOS, WMO GAW, NOAA flask network, AGAGE).

L320ff: Again, a reference to existing observing networks that are used for source attribution would make a lot of sense.

L324: I suppose with 'basic back-trajectory analysis' you are referring to what I called the fourth (non-statistical) method, that only looks at trajectories/footprints of individual observations. I don't quite understand how such an approach is reflected in Fig. 6.

L362f: Is the selection of one observation every two days enforced onto the already cropped, one-year data set or on the two-year data set? If the latter, then it is surprising that a decrease in performance is observed, whereas none was seen when shortening the time series from two to one year. Both cases should have the same number of total observations in the analysis.

L364: This is a trivial conclusion. More observations should always improve this kind of statistical source attribution analysis. However, there is also a limit to enhancing temporal resolution, since atmospheric variables are usually auto-correlated and the amount of independent information cannot be increased by measuring more frequently. However, this is for time scales shorter than a day and is probably not what was referred to here.

L370: Consider 'arbitrary results' instead of 'insignificant information'.

L373f: Exactly, that's why filtering for a remote source may be dangerous and the numbers obtained here cannot easily be generalised for other sites or compounds.

L407: Gilardoni et al. (2019). There is no source attribution analysis presented in this document.

L422: Panel a of Fig. 9 does not show a ratio map.

L427: How does the potentially long transport time from the Caspian Sea agree with atmospheric lifetimes of MSA? Would we not expect MSA to be mostly destroyed?

L468: Repeated from comment above. It's probably more the shadowing effect than the continental dominance.

L501: 'standard ratio method is more advanced than widely used inverse modeling analysis'. This is certainly not true, unless you wanted to express that ratio methods were the most frequently applied tool for analysing INP sources in the Arctic.

Technical comments

L27: '.' missing after (Matus and L'Ecuyer, 2017).

The bibliography does not comply with the Copernicus style.

References

Brunner, D., Henne, S., Keller, C. A., Vollmer, M. K., and Reimann, S.: Estimating European Halocarbon Emissions Using Lagrangian Backward Transport Modeling and in Situ Measurements at the Jungfraujoch High-Alpine Site, in: Lagrangian Modeling of the Atmosphere, edited by: Lin, J. C., Gerbig, C., Brunner, D., Stohl, A., Luhar, A., and Webley, P., Geophysical Monographs, AGU, Washington, DC, 207-221, doi: 10.1029/2012gm001258, 2013.

Fang, X., Saito, T., Park, S., Li, S., Yokouchi, Y., and Prinn, R. G.: Performance of Back-Trajectory Statistical Methods and Inverse Modeling Method in Locating Emission Sources, ACS Earth and Space Chemistry, 2, 843-851, doi: 10.1029/2012gm001258, 2018.

Jones, A., Thomson, D., Hort, M., and Devenish, B.: The U.K. Met Office's Next-Generation Atmospheric Dispersion Model, NAME III, Boston, MA, 2007, 10.1007/978-0-387-68854-1_62, 580-589, doi: 10.1007/978-0-387-68854-1_62, 2007.

Resovsky, A., Ramonet, M., Rivier, L., Tarniewicz, J., Ciais, P., Steinbacher, M., Mammarella, I., Mölder, M., Heliasz, M., Kubistin, D., Lindauer, M., Müller-Williams, J., Conil, S., and Engelen, R.: An algorithm to detect non-background signals in greenhouse gas time series from European tall tower and mountain stations, Atmos. Meas. Tech., 14, 6119-6135, doi: 10.5194/amt-14-6119-2021, 2021.

Ruckstuhl, A. F., Henne, S., Reimann, S., Steinbacher, M., Vollmer, M. K., O'Doherty, S., Buchmann, B., and Hueglin, C.: Robust extraction of baseline signal of atmospheric trace species using local regression, Atmos. Meas. Tech., 5, 2613-2624, doi: 10.5194/amt-5-2613-2012, 2012.