| 18 Feb 2026
Limitations of 2-pool models in representing different time-scale dynamics of particulate and mineral-associated organic carbon
Abstract. In the last decade, the conceptual framework that characterizes soil organic carbon (SOC) into particulate organic carbon (POC) and mineral-associated organic carbon (MAOC) fractions has gained traction in studies of C dynamics. This SOC characterization is useful for developing empirical studies and for parsimonious model parameterizations. However, rigorous testing of model structures incorporating the POC-MAOC framework is still lacking, in particular tests involving simultaneous measurements of C pool changes and respiration fluxes. We conducted an incubation experiment using control and litter-addition treatments, measuring changes in SOC fraction contents and respiration fluxes throughout the incubation. Then, we applied an inverse modelling approach to compare the performance of 2-pool (POC-MAOC) and 3-pool models (which also included a faster-cycling litter C pool) to reproduce the observed data. We then calculated the C ages and transit times to explore the predicted C persistence. Finally, we performed simulations to evaluate the effects of different model structures and parameterizations on SOC persistence. For both treatments, we observed that 2-pool models were unable to simultaneously reproduce the changes in C pool contents and respiration, while the 3-pool models adequately predicted both variables and yielded lower C ages and transit times. The fact that 3-pool models outperformed 2-pool models even for control soils, indicates that POC represents a heterogeneous pool that should be modelled as distinct compartments. We discuss that 2-pool models collapse POC dynamics operating at different timescales into a single one, failing to capture the different respiration phases and the gradual C pool changes. In contrast, 3-pool models distributed C processes operating at different timescales among compartments: the litter C pool captured faster-cycling dynamics, allowing POC and MAOC to better represent intermediate- and long-term dynamics, respectively. We also found that both model structure and changes in key parameters affected C persistence estimations. Models that included shorter pathways to MAOC, or allowed faster transfers into more persistent pools, predicted higher C age and transit time. This study highlights the limitations of representing SOC dynamics exclusively through POC and MAOC and shows how model structure shapes SOC contents and persistence estimates. Rather than advocating a specific model configuration, our results suggest that SOC models should explicitly represent processes operating across multiple timescales, which, depending on the ecosystem context, may require incorporating additional C compartments beyond the POC-MAOC framework. Furthermore, as transfer rates play a key role in determining SOC persistence, it is important to better understand and quantify how C is transferred toward MAOC and how these processes can be represented in models.
The manuscript submitted by Fernandez-Catinot et al. combines an incubation experiment with SOC fractionation and regular respiration measurements to test whether 2-pool POC-MAOC models can effectively capture both long-term C pool dynamics and short-term respiration fluxes. They find that a third, fast-cycling compartment is needed to accurately capture both simultaneously. The subsequent simulation study shows that transfer rates into MAOC, rather than litter decomposition rates, are the primary driver of predicted C persistence, and that model structure and parameter changes interact non-additively, producing up to 7-fold differences in estimated transit time. The experiment is well designed, and the results are an important contribution to discussions around SOC modeling methodologies.
While I have no major concerns, there are several issues that I believe should be addressed in revision. Most importantly, the control soil comparison rests on an important assumption about the initial Litter C fraction which is currently untested. I would like to see a sensitivity analysis or a discussion of how this choice affects the results. I also encourage the authors to acknowledge the constraints that a 6-month incubation places on steady-state C persistence estimates, and to consider cross-validating fitted parameters between treatments. I provide a few additional comments below.
L53: "mainly produced by microbial re-synthesis" — consider softening; direct sorption of plant-derived compounds to mineral surfaces is also an important MAOC formation pathway (e.g., Whalen et al. 2022)
L72: "despite its components might cycle" — grammar; should be "despite the fact that its components might cycle" or similar
L90 (and throughout): When referring to specific numerical comparisons, "mean transit time" and "mean C age" would be more precise, since the full metrics are distributions. This applies here and in several other places (e.g., L28, L35, L334, L353, L397).
L125-132: Please report mass recovery and C recovery for fractions derived from the wet sieving fractionation.
L152-154: It is unclear whether the a_ij values reported in Table 2 are actually absolute transfer rates (yr-1) as the footnote states. All reported values fall between 0 and 1, and under the absolute rate interpretation some parameter sets would violate the non-negative respiration constraint (e.g., control 3-pool: k2 = 0.121 but a32 = 0.968), suggesting the values may be proportions. Please check and correct as needed.
L156-160: For control soils, the 3-pool model assigns 10% of POC to a fast-cycling Litter C pool, in line with the author's argument that the POC pool is heterogeneous. However, for the litter-addition treatment, the Litter C pool appears to contain only the added litter, with the native fast-cycling fraction folded back into POC. If POC heterogeneity is a general property of these soils, the litter-addition model's Litter C pool should include both added and native fast-cycling material. I recommend the authors should clarify this choice and discuss whether it affects the fitted parameters or model performance.
L160: Additionally, for control soils, the 3-pool model requires assigning a fraction of initial POC to the Litter C pool. In the manuscript, the authors choose an arbitrary value of 10% (L160) but do not report how sensitive the results are to this choice. Since the litter-addition result is acknowledged as "almost self-evident" (L321), the control soil comparison is the stronger, non-trivial test of the 3-pool model's advantage, however, it depends in large part on this assumption. I recommend that the authors either (a) report the sensitivity of the 3-pool model's advantage/parameterization to changes in the Litter C fractions (e.g., 5-20%), or (b) attempt to fit the initial Litter C fraction as an additional parameter in the inverse optimization (if identifiability permits).
L260-267: Given that the C age and transit time estimates are derived from steady-state equations applied to parameters fitted from a 6-month incubation, over which the MAOC pool barely changes, I think the k_MAOC may be poorly constrained. Since the steady-state persistence metrics shown in Figure 4 are sensitive to the parameters of the slowest pool, I am concerned that the C age and transit time estimates may carry important uncertainty. I'd recommend the authors report confidence intervals on the fitted parameters and derived metrics to help address this concern. Additionally, the input vector u used to calculate transit time (Eq. 7) should be specified, given that I = 0 during the incubation (L145).
L342: Also see Curtin et al. 2019 (https://doi.org/10.1016/S1002-0160(18)60049-9) on POM as an organo-mineral composite with physically distinct sub-fractions.
L345-351: The Litter C pool in control soils is a modeling construct that cannot be independently measured through the fractionation used here — chopped litter and native POC both land in the >53 um fraction. I think this is worth acknowledging, as it means the 3-pool model's advantage for control soils cannot be validated against measured pool sizes at any time point, similar to the pools of classic process based models like CENTURY.
Section 4.1: Since the 2- and 3-pool models are fitted independently to each treatment, I'm curious whether the 3-pool parameters fitted to the litter-addition treatment would also predict the control soil dynamics (or vice versa), with only the Litter C initial condition changing. If so, that would be strong evidence that the 3-pool structure captures a real property of these soils rather than reflecting additional fitting flexibility. This may be beyond the scope of the current revision, but it could act as a useful cross validation and I encourage the authors to consider it.