General comments

The authors present a new interesting tool, AgPaDS, developed to rapidly simulate Lagrangian transport in the atmosphere, coupled with advanced visualization and interactive simulation configurations,  for agricultural applications. They compare its performance with  HYSPLIT, one of the mostly used atmospheric transport models, demonstrating a significant increase in computational efficiency across different tests, thanks to the use of GPU.

This tool goes into the direction of solving persistent challenges faced by agricultural researchers, such as (a) the simulation time (which is dealt by directly in the model) and (b) the coupling of windborne pathogen transport with plant epidemiological dynamic models (which the authors propose as a possible advancement for this tool).

Specific comments

Undoubtedly, the improvement in computation time compared to the benchmark is outstanding, and the visualization tools are also remarkable. The article’s focus is more on the algorithmic aspects than on the Lagrangian implementations, for which the authors drew inspiration from NAME. I recognize that this is a very technical paper, and as someone not deeply familiar with Lagrangian modeling at this level of implementation, I admit it is not always easy to follow. For instance, (i) I would have preferred the model description (section 3) to appear first in the Materials and Methods; (ii) acronyms (GPU, ECMWF, CUDA) are used without explanation. On the other hand, despite the claims of a better compatibility of this tool with crop or epidemiological models, the coupling does not seem straightforward.

I commend the authors for their accurate model evaluation setups, both in terms of experimental design and in terms of indicators used to measure the comparisons. However, I believe that a couple of points require revision or at least further discussion. In the second experiment (Table 3), the authors state that differences between HYSPLIT simulations and AgPaDS are less than one order of magnitude, but the data actually show a one-order-of-magnitude difference; this is not inadequate per se, but needs to be better contextualized. Moreover, in the third set of experiments (~Line 735), when comparing the simulation of atmospheric transport of Phakopsora spores by hurricane Ivan if it would be feasible to compare the simulated and observed deposition/presence of the soybean rust infections in USA (and not on the severity), instead of leaving it qualitatively.

My last question is, given the strong inspiration drawn from NAME, is there a specific reason why the authors chose to benchmark against HYSPLIT rather than on NAME?

Minor corrections

• L 15: GPU, which is the Graphical Processing Unit.
• 38-39 “et al.” is sometimes italicized, but not always. Please homogenize throughout the text.
• 40. So far, the introduction quotes very general papers, without going into details. For instance, the authors do not tell the name of any of these “devastating crop diseases” (L. 38) and seem to suggest that long-distance-dispersed plant pathogens cause 17-30% of yield loss (L.40), while these percentages aggregate all infectious plant diseases.
• 52: atmospheric transport models or Atmospheric Transport Models (as in L. 10)?
• 80: What does I/O mean? Please, define acronyms throughout the text.
• 85: “These challenges evolve around, unknowns around pathogen viability decay during atmospheric transport: uncertainty estimates for processes involved in atmospheric transmission of crop pathogens”. What does it mean?
• L81-132: these bullet points are interesting and cover exhaustively the issues with ATMs, but it think they could gain in readability – for example, by splitting or shortening some sentences and capitalizing the initial letter of each point.
• 97: I wonder if the authors meant “links” and not “vertices” (which I assume is asynonym of “nodes”).
• L 120: Missing full stop.
• 140: Is the figure correctly placed here?
• 210: P in Python should be capitalized, here and in the rest of the text
• 219: Specify what ECMWF is.
• 361-369: What do subscripts v, p, lambda and phi stand for? Also, is m_v the same as m_i (I do not think so)?
• 394: SSD?
• 397: CUDA? Quoted 33 times, also in Fig. 2
• 460-461: I am not sure of the rationale. Is N the number of simulation particles as in L. 371?
• 490: Consider telling more explicitly the reader the dimensions of wind data in ERA 5 (x, y, pression), not just as unit of measures.
• 503: Consider telling the reader what a timescale is.
• 510: no space before comma.
• 537-554: A very interesting overview of possible techniques of implementing viability decay!
• 570: This is not the first time I see this value of this parameter used for wet deposition. This is quite evidently computed on an approximation, since 25.4 mm = 1 inch and 63.2% is just 100*(1 – exp(-1)), or the expected loss after 1 unit of time. It is a pity that there are better estimations for this parameter.
• 675: Despite atmospheric trajectories and material deposition look very similar, the same can not be said about Lagrangian particles in the air. Why?
• 692: Also relative differences HYSPLIT – AgPaDS look quite important int Table 2. Could you please discuss?
• 710: The same can be said of Table 3, for example of the median deposition value. Despite what you state in L. 712, thz difference between 5.8* 10^-10 and 2.8* 10^-11 is ~1 order of magnitude.
• 735: “Whilst the available data does not allow for exact quantitative evaluation”. This sentence makes me ask if (1) there were no better episodes of LLDto test the model performances, such as stem rust of wheat in East Africa or Asia or (2) a test to compare the presence (and not on the severity) of the soybean rust infections in the USA and your simulations would be feasible.
• Line 873: “in comprehensive”.

This manuscript presents AgPaDS, a GPU‑accelerated Lagrangian atmospheric transport model with interactive 3‑D in situ visualisation, motivated primarily by applications in crop epidemiology. The paper describes in exceptional technical detail the software architecture, GPU implementation, and visualisation capabilities, and demonstrates that the model reproduces results broadly consistent with established atmospheric transport models (HYSPLIT, IAMS) while achieving substantial performance gains. This is a strong and valuable model description paper but I have some concerns regarding balance, framing, and evaluation that should be addressed prior to publication.

Major issues

1. The work represents a significant engineering and visualisation achievement and fulfills many criteria of a GMD model description paper. However, while the manuscript is positioned as a model description paper, a substantial fraction of the text focuses on GPU architecture, CUDA/OpenGL implementation, GUI design, and visualisation pipelines, often at greater length and detail than the atmospheric or biological modelling itself. Hence, the manuscript would benefit from clearer framing of its scientific contribution, greater balance between software engineering and atmospheric/biological modelling, and a more critical discussion of limitations and use‑case constraints because the scientific novelty risks being overshadowed by the engineering narrative. Consider shortening the discussion of enabling technology and reallocating space to sensitivity of results to modelling assumptions, limitations and failure modes, and scientific implications for crop epidemiology beyond “speed + visualisation”.
2. Model evaluation is largely based on qualitative visual agreement, relative comparison to HYSPLIT and IAMS, with broad consistency against empirical soybean rust observations. While this is reasonable given limited data, agreement with another model does not constitute validation, and some reported differences (e.g. 24% longer mean trajectories; 34-43% differences in plume spread) are non‑negligible for epidemiological applications. The evaluation results should be clearly framed as benchmarking against existing models, not validation. Also consider strengthening the discussion of which deviations matter (and for which questions), which applications are robust to observed discrepancies, which applications require caution or additional calibration. I also recommend adding idealised test cases (e.g. homogeneous flow, no turbulence) where expected behaviour is conceptually clear and demonstrated.
3. Epidemiological relevance is aspirational rather than demonstrated because the manuscript does not present a clear end‑to‑end epidemiological analysis or a decision‑relevant use case. This means that most epidemiological benefits are presented as future potential, rather than demonstrated outcomes of the new modelling framework. No crop‑disease risk metric is produced, no comparison to observed disease timing/intensity beyond spatial coincidence is provided, and no assessment is made of how interactive visualisation could change decisions. I recommend that the authors either add one concrete demonstration of epidemiological utility or more explicitly limit and qualify claims regarding operational relevance and early‑warning systems.
4. The atmospheric transport formulation closely follows established Lagrangian frameworks, and many parameterisation schemes are largely adopted from earlier work. The rationale and sensitivity for those parameter values, and the interaction with GPU performance constraints are not fully discussed. A short sensitivity analysis or explicit discussion of which parameters are intended to be exploratory versus physically constrained would substantially strengthen the model description.
5. While I am not a specialist in high‑performance computing, the validity of the computational speed comparison to HYSPLIT is not entirely clear, as performance is evaluated against a single‑CPU configuration rather than commonly used parallel CPU setups. Whilst this fact is acknowledged, it should also be stated that such configurations would reduce the apparent advantage of AgPaDS. Clarifying this limitation explicitly would strengthen the transparency of the performance assessment rather than weaken the contribution. In addition, while the manuscript draws conceptually on NAME (e.g. for turbulence parameterisation), no explicit comparison to NAME is presented. This is not necessarily required, but the absence should be more clearly acknowledged and its implications discussed, particularly in the context of performance claims and plume spread differences.

Minor issues

1. The manuscript is very long and there are several sections that repeat the same points, particularly in the Introduction and the Discussion. Careful revision of long sentences and the removal of duplicated material could shorten the manuscript overall.
2. Some symbols and notations are introduced before being defined.
3. Variable naming is not always consistent, e.g. mi is used to define the material carried by each simulation particle in equation 3, but pathogen viability in equations 11 and 12, and mv is used to define pathogen viability decay in equation 1, and sometimes m and M are used to define particle mass.
4. Tables 2 and 3 are information‑ Highlighting key numbers (e.g. in bold) would help.
5. L22: The sentence “We show that AgPaDS maintains good agreement with HYSPLIT” is overstating validation, suggest changing to something like “We show that AgPaDS produces atmospheric transport patterns broadly consistent with HYSPLIT across a set of benchmark cases”
6. L697-698: The sentence “moderately longer mean distances from the source (24%), compared with HYSPLIT” should qualify the implications, e.g. “which may be consequential for applications sensitive to arrival timing or long‑range thresholds”
7. The figure on page 40 of the supplement is incorrectly labelled Figure S3 instead of S35.
8. The figure on page 32 of the supplement is incorrectly labelled Figure S272 instead of S27.
9. Comparison is misspelt in the caption for Figure S29 in the supplement.
10. Marocco should be Morocco in the supplement
11. The caption for Figure S3 in the supplement contains duplicate “(ii)”.
12. Many of the figures in the supplement are very small and the text is difficult to read
13. The supplement table of contents lists figures and tables by their page numbers, but no page numbers are provided in the supplement.