This manuscript presents a combination of CAMP’s support for multiphase chemistry with TChem’s computational model, based on the Kokkos library, to enable gaseous and aerosol chemistry on CPU-only and heterogeneous

architectures. User flexibility is provided through the support of multiple ODE solvers.

The authors demonstrate integration with the PartMC particle-resolved model and validate their results against the previous PartMC–CAMP implementation. The paper is generally well-structured and clearly articulates the need for

this work in the context of improving the performance and architecture compatibility of atmospheric chemistry-aerosol models. 

The documented code development represents a step toward performance-portable chemical solvers for next-generation atmospheric models, and is timely and relevant to the scope and audience of the GMD journal. 

The methodology presented in Sections 1 and 2 is sound, innovative, and necessary for the next generation of atmospheric models. The core idea of combining CAMP's multiphase flexibility with TChem/Kokkos performance portability is a solid contribution. However, the manuscript, while strong on methodology and capability, is currently limited in its demonstration of performance and results for real-world applications.

General Comments:

The code can readily support sectional and per-particle models.  In principle,  TChem-atm can be adapted to support more mechanisms than the two curently supported (UCI, CB05). However, it should be made clear (also in the abstract) that modal host models need to be adapted to support TChem-atm, and that architecute-specific performance tuning is required (for example a factor of 2 improvement is reported for the NVIDIA H100) to achieve higher efficiency.

The current validation focuses on a box-model scenario using the CB05 mechanism and SIMPOL partitioning.

While this is adequate as a proof-of-concept, it falls short of a scientifically realistic demonstration (e.g., regional-scale or global model simulation) that would better showcase TChem-atm’s capabilities and its potential to speed-up atmospheric chemistry-aerosol production simulations. Again, in the abstract (and elsewhere) it would be more fitting to describe

the results presented here as "proof-of-concept".

If possible, the manuscript could benefit from a brief profiling summary (e.g., fraction of time spent in kernel execution vs. memory transfer) to strengthen the discussion. 

Finally, it is not clear why the authors evaluated the performance of direct linear solvers for dense (full) Jacobians for atmospheric chemistrykinetics (where Jacobians are generally sparse, as is the case for the mechanism tested here). Thus, the reported results in terms of memory limitation, and longer time-to-solution are to be expected and almost trivial.

And of course, preconditioned iterative methods are preferrable for the large, stiff ODE systems of atmospheric chemical kinetics because they require significantly less memory than direct solvers.  It is suggeted to refactor the disucssion around this.

Specific Comments:

The reaction mechanisms in Fig.3 do not appear correctly in the preprint. Please check.

Line 136 and elsewhere: Please be consistent with capitalisation and code names e.g. "Sundials CVODE" or "SUNDIALS", etc.

Sec. 2.5 only discusses the partitioning of semi-volatile species. It's not clear how other cases of heterogeneous reactions are handled.

Fig. 5: It's clear that there's agreement as also reported in the text, so this figure offers little for the reader.  Please consider plotting the differences if possible between runs instead, to reveal more architecture-related information.

 Line 326: The reference for SIMPOL is wrong. Please correct. 

This publication describes the implementation of TChem-atm (v2.0.0), a software designed to permit efficient resolution of of stiff ODEs of atmospheric chemistry, including interaction between gas-phase and aerosol-phase compounds, using Kokkos to avoid being dependant to specific hardware-dependant optimizations in the code. This publication is timely, and relevant to the scope of GMD. I believe that there are a few shortcomings in the study, in particular the two points I list as general comments : clarify the relationship between TChem-atm, and earlier components TChem and CAMP. In this respect, I suggest adding a figure showing how things are organised within TChem-atm, and possible inclusions from other models. I also suggest that the authors explain how they see the road towards the use of TChem in real-world geophysical applications (if possible, based on previous work with TChem and CAMP in this field).

Even though I do not think that these shortcomings question the relevance of the article or the quality of the results, I think the lack of clarity regarding these aspects requires major revisions to make sure before publication that the paper contains all the useful information for the geophysical modelling community.



Please see the details of my review in the attached document.