| 30 Apr 2026
GENOA v3: A flexible framework for reduction and exploration of highly detailed chemical mechanisms
Abstract. Comprehensive atmospheric chemical mechanisms for volatile organic compound (VOC) oxidation contain thousands to millions of reactions and species, presenting major computational challenges for large-scale or long-term simulations. As mechanism complexity continues to increase, reduction strategies are required to enable their use in atmospheric modeling while preserving accuracy.
This paper presents the GENerator of Optimized Atmospheric chemical mechanisms (GENOA v3), a major advancement over earlier GENOA versions that enables scalable reduction of highly detailed mechanisms containing up to millions of reactions and species. GENOA v3 combines fast, strategy-driven threshold-based reduction (TBR) with simulation-based reduction (SBR) that explicitly controls accuracy. The framework is modular, graph-aware, and user-configurable, resulting in compact and chemically interpretable reduced mechanisms.
Applications to GECKO-A mechanisms for diverse VOC precursors across a range of scenarios show that TBR achieves mechanism size reductions of 20–90 % while preserving reasonable accuracy for metrics related to secondary organic aerosol (SOA) formation and gas-phase chemistry, with performance systematically dependent on precursor structure and chemical complexity across mechanisms. SBR achieves further reductions in mechanism size by several orders of magnitude; when trained with 15 % mean error constraints, SBR produces schemes within 0.02 % of the original size for preservation of SOA mass and 0.05 % when also considering gas-phase reactivity (e.g., OH, O3, and NO3).
These results demonstrate for the first time that GENOA v3 can reduce highly detailed chemical mechanisms while jointly preserving SOA mass and gas-phase reactivity, achieving substantial size reductions with reasonable accuracy across a wide range of scenarios. Continued application of GENOA v3 and growth of a user community will potentially support the development of libraries of reduced mechanisms and optimized reduction strategy sets tailored to specific modeling applications.
The paper presents the GENOA v3 framework for reduction of atmospheric chemical mechanisms with huge number of species and reactions such as the highly detailed VOC reaction mechanisms generated by GECKO-A. The work demonstrates that GENOA v3 can effectively reduce highly detailed VOC mechanisms while preserving SOA mass production and gas-phase reactivity with reasonable accuracy across a wide range of atmospherically relevant scenarios.
The reduction follows a rule-based strategy and is performed in two subsequent approaches, first through a threshold-based reduction (TBR) and second through a simulation-based reduction (SBR) that uses the TBR-reduced mechanism as a starting point. Whereas TBR provides a rapid initial pruning of large mechanisms at low computational demand, only the subsequent SBR provides the aggressive reduction needed for application in a large-scale CTM. It is not apparent whether the TBR part can be skipped in principle or whether it is required for reordering or attenuating the starting-point mechanism before it can be used in SBR. The authors currently imply that the starting mechanism needs some refinement before it is possible to run it efficiently in SBR. However, GECKO-A should already generate a mechanism that is not too stiff for theoretically applying it in a 0-D box model.
SBR is not a black box but the reduction comes with the risk that the reduced mechanism does not remain accurate under all given or even not the most critical atmospheric conditions. The error control during mechanism reduction is somewhat vague. Are there any checks and controls happening during the reduction process chain such as atom balance (loss of C or N or O)? The user-defined error tolerances for the reduction errors in SBR are also part of this concern. Not only does the choice of the error tolerance (ε) increase the degrees of freedom of a reduced mechanism created from the same starting mechanism by different users, it is also not demonstrated in this work whether the error of the reduced mechanism is related to the value of ε. Therefore, it is not guaranteed that decreasing ε results in a monotonically decreasing error of simulated concentrations compared to the reference, and may not always provide the smallest reduced mechanism at a given simulation error. Probably different epsilon values should be tested by the user and could result in a reduced mechanism with less species but only slightly different simulation error in comparison to the reference mechanism. My suggestion is to study how number of species and simulation error (NMB) change as function of epsilon for a subset of the mechanism, for example U8.
The more compact chemical mechanism resulting from GENOA v3 may be viable for large-scale atmospheric simulations, although this is not demonstrated in the paper. Since each user-defined set of configurations of targets and error tolerances results in a differently reduced mechanism, it will be difficult for those users who apply the reduced mechanism in a CTM simulation to define the mechanism. The use of the GENOA-reduced mechanism in CTM simulations would require the publication of the full reduced mechanism including the user-specific reduction protocol either before or along with the CTM simulation when publishing model results. A nomenclature for the obtained reduced chemical mechanism should be suggested in this work.
The evaluation of the SBR reduction performance (section 6) is well organized and presented. The reaction pathway visualization is excellent and the example of reaction pathways for U8 with SOA-focused reductions at high and medium error tolerances versus multi-target reduction is well chosen. SBR is the main development of this work, and this part should be more in focus of the paper. Given the minor role of TBR in reducing the original mechanism, it would be sufficient to describe TBR in form of a technical manual in the Supplement or at least to shorten the TBR evaluation part in the manuscript.
The development of GENOA v3 is a very useful instrument for atmospheric chemistry modelling. The evaluation of reduction candidates in terms of SOA and multiple gas-phase reactivity metrices and the capacity to reduce starting mechanisms containing hundred thousands of reactions into compact mechanisms makes this work a valuable contribution.
The revision of the manuscript should include a comparison of the GENOA-reduced mechanism against other condensed chemistry mechanisms such as the Common Representative Intermediate (CRI) scheme from reducing MCM. In addition, my concerns listed below should be addressed.
Specific Comments:
1.) Introduction: two categories of systematic mechanism reduction are put forth in the introduction: chemistry informed rule-based reduction versus machine learning/data driven reduction. The agglomeration of all non-ML techniques into one category is a simplification that ignores other possibilities for mathematical chemistry-based reductions. In the field of combustion flame kinetics, rigorous mechanism reduction has been developed much earlier than in the field of atmospheric chemistry. Nagy and Turányi (2009) developed the simulation error minimization connectivity method which generates thousands of candidate reduced mechanisms on the basis of the inspection of the normalized Jacobian and produces a database, in which the errors of simulations using the reduced mechanisms are recorded. Based on this database, an optimal reduced mechanism is selected. The directed relation graph (DRG) method uses a directed graph to map the coupling of specues and consequently find unimportant species for removal based on an acceptable threshold. Sander et al. (2019) have used the graph-oriented reduction method with error propagation by Niemeyer et al. (2010) to reduce the Mainz Organic Mechanism (MOM) and compared the reduced and the full mechanism in a global simulation. For completeness, MOM that is included in the CAABA/ MECCA box model (Sander etal., 2019) should be mentioned as another representative example of detailed VOC mechanisms.
2.) Section 2.1: before introducing the highlights of GENOA v3, authors should put forth what the criteria are for measuring the performance of GENOA in this study, in terms of scalability, simulation errors in tSOA, simulation errors in O3 production, mass balance, atom balance of C, O and N, etc.
3.) Figure 2: Does the gray downwards arrow mean that output of GENOA v1 and v2 can be used as input for GENOA v3?
4.) Section 3.2.2, branch concentration removal: While it makes sense to remove minor branches, a description is missing how the remaining relative branching ratios are corrected after removing the unimportant branch.
5.) Section 3.3, introduction of SBR: It is stated that the candidate (for the same portion of the mechanism) is selected that achieves the best balance between reduction efficiency and accuracy. How is it ensured that the selected candidate is the best candidate for reduction overall? In other words, how can we be sure that the selected candidate and its subsequent further reduction preclude that a less favorable candidate turn out as better reduction option a few stages downstream? What are “user-defined stopping conditions” and where are they specified?
6.) Section 3.3.3, candidate evaluation criteria: elaborate more on the error tolerances. It seems odd to let the user define the error tolerances. Assuming that the user-defined error tolerances are carefully documented in each reduction step, this nevertheless imposes knowledge about prescription of error tolerances on the users. Please describe the objective reasoning for the specific error tolerances (also see my suggestion in the general comment).
7.) Section 4.1, test mechanism selection: Include a reference to Table 6 already here. Consider including the total number of RO2 species in Table 6.
8.) P16, L 410: Introduce briefly which SOA formation algorithm is applied in the box model simulation. Is the initial SOA seed mass non-volatile? Is the partitioning reversible?
9.) Figure 3: why is there no 10% limit shown in plot (b)?
10.) Table 5: Explain in the text, why the error tolerances increase within one stage and throughout the stages. P22, L 501: It is mentioned here that error tolerances are relaxed after stage A2 without giving the rationale behind this. P22, L 511: Did you test how sensitive is the reduced mechanism to mean reduction errors larger than 10%, for example in the prediction of tSOA, and how does the choice of larger reduction errors affect the size reduction?
11.) P24, L 550: What are the criteria to consider a tSOA error in TBR as acceptable? That should be defined from the outset and may also be different for different PVOC. Related to this (P26, L 583), it is shown that tSOA in T10 is systematically giving 10-20% lower SOA mass than the Reference. This potentially leads to a systematic underestimation of observed SOA in simulation with a box model using this reaction mechanism. Please comment on that issue and how it can be circumvented.
12.) P31, L679-686: A possible conclusion at the end of the TBR performance evaluation could be that T6 can be used for multitarget reduction (gas-phase and SOA related) and T10 only for SOA-related reductions. This also could lead to the conclusion that it is better to streamline reductions with TBR by just allowing two different options, T6 or T10, with a given set of pre-defined reduction parameters for each.
13.) Section 7.2.1, scenario dependence: Not spanning scenarios that occur under certain atmospheric conditions put the mechanism reduction at risk, although the selected 11 scenarios cover most typical environments in the lower atmosphere. An additional scenario should include light-dependent emissions of isoprene, as mixing with isoprene may potentially suppress SOA and this interference is important everywhere in the continental atmosphere. Moreover, a scenario representing conditions of the upper troposphere in the tropics would assist in the reduction for use in global chemistry-transport models.
14.) Section 7.4, integration with machine learning: Although GENOA retains mechanistic interpretability, neither ML-based approaches nor GENOA allow tracking of reduction decisions to find out why a certain reduction candidate was preferred over others. With GENOA there is at least theoretically a chance to trace a reduction decision, but I guess it is terribly complicated to trace it back after several reduction stages. Please comment.
References:
Nagy, T. and Turanyi, T.: Reduction of very large reaction mechanisms using methods based on simulation error minimization, Combustion and Flame, 156, 417–428, https://doi.org/10.1016/j.combustflame.2008.11.001, 2009.
Niemeyer, K. E., Sung, C.-J., and Raju, M. P.: Skeletal mechanism generation for surrogate fuels using directed relation graph with error propagation and sensitivity analysis, Combustion and Flame, 157, 1760–1770, https://doi.org/10.1016/j.combustflame.2009.12.022, 2010.
Sander, R., Baumgaertner, A., Cabrera-Perez, D., Frank, F., Gromov, S., Grooß, J.-U., Harder, H., Huijnen, V., Jöckel, P., Karydis, V. A., Niemeyer, K. E., Pozzer, A., Riede, H., Schultz, M. G., Taraborrelli, D., and Tauer, S.: The community atmospheric chemistry box model CAABA/MECCA-4.0, Geosci. Model Dev., 12, 1365–1385, https://doi.org/10.5194/gmd-12-1365-2019, 2019.