see attached file

    In this article, the authors proposed one novel algorithm --NeuPlume, which has the potentiial to address a long-standing difficult and relevant problem: estimating point-source emissions and reconstructing a three-dimensional concentration field from sparse observations. The manuscript is clearly organized, and the combination of synthetic tests, UAV measurements, ablation experiments, and boundary diagnostics gives the study a solid starting point. I also appreciate the care taken to state the present domain of application, including neutral conditions, flat terrain, low source heights, and a single point source. The discussion of open boundaries, wind-speed representativeness, and computational cost is unusually candid for a methods paper.

     I find the framework promising and well within the scope of Atmospheric Chemistry and Physics. At the same time, several claims about probabilistic inference and method performance need firmer support. I therefore recommend minor revision. The points below are intended to help the authors clarify what NeuPlume currently establishes and make the strongest parts of the study easier to assess.

 Main comments:

1. What probability distribution is being estimated?

   The manuscript does not yet show how the temperature relates to the stated observation noise, residual scale, number of observations, parameter prior, or grid-cell volume.  Please define the observation model, parameter prior, grid measure, loss function, and temperature in one consistent account. The manuscript should also explain what source of conditional field variability is represented in the training data. If the Gibbs weights are intended as an empirical approximation, the terminology throughout the abstract, methods, results, and conclusions should reflect that definition, and the intervals should be described accordingly.

 2. Isolate the contribution of diffusion and DPS

     I suggest comparing at least (i) the Hash-INR or conditional surrogate with the same coarse-to-fine search but no diffusion model, and (ii) conditional diffusion with the DPS guidance removed. These variants should use the same observations, parameter grid, loss, random seeds, and computational budget. Parameter error alone would give an incomplete picture; field-reconstruction error, uncertainty behavior, and runtime should be reported as well. The resulting comparison would allow the authors to state plainly what diffusion adds and whether DPS contributes to accuracy, field diversity, uncertainty estimation, or some other part of the workflow.

    3. The baseline comparison mixes method performance with model and sampling mismatch

     The comparison would be more informative if each method were evaluated under a suitable and comparable observation geometry. A closer baseline would also help, for example a conditional neural surrogate coupled to a conventional grid or Bayesian inversion. Such a test would separate the benefit of the forward surrogate from the benefit of the generative component. If these additional comparisons are outside the revision scope, the claims should be tied explicitly to the LSM-generated test fields, parameter ranges, and random volumetric sampling protocol used here. In particular, the current results do not yet support broad statements that the CNF itself supplies stronger physical constraints than simplified dispersion models.

 

4. Bring the field-reconstruction evidence into the main paper

    Full-field recovery is one of the most interesting features of NeuPlume, yet much of the quantitative support appears in the Supplement. The main paper should include a representative independent test case with a consistent reconstruction metric. Figures 2 and 3 would also benefit from complete coordinate ticks, units, an unambiguous color scale, and a compact numerical indication of reconstruction error. The caption should explain the source of pixelation or particle noise. Visual texture by itself is not sufficient evidence of non-Gaussian plume morphology.

 5. Report uncertainty around the calibration estimates

     Please report the numerator, denominator, coverage estimate, and a binomial confidence interval for each parameter at each nominal level. The data used to choose the temperature should be identified separately from the data used to assess coverage. Figure 7, its caption, and the accompanying text should show the same parameter set and nominal levels. A short discussion of the limited sample size would make the calibration result much easier to interpret.

 6. Draw out the atmospheric measurement implications

    The manuscript currently devotes more attention to model components than to what the experiments reveal about measurement design. For ACP readers, the diagnostics on observation coverage, distance from the source, wind-speed bias, and boundary losses may be at least as useful as the architecture itself. These findings should be synthesized into a concise, conditional account of when the source parameters are identifiable under the tested setup.

     I encourage the authors to discuss which observation geometries and transport mismatches most strongly affected the inversion, and how those findings could inform future UAV sampling or controlled-release experiments. The limits of the evidence should remain explicit: the approximate 1% coverage and 250 m distance results apply to the domain, sampling protocol, and forward model examined in this study. Framed this way, the diagnostics can offer practical guidance without being presented as universal thresholds.