The impacts of model structure, parameter uncertainty and experimental design on Earth system model simulations of litter bag decomposition experiments

Note: This manuscript has been authored by UT-Battelle, LLC under Contract No. DE-AC05-00OR22725 with the U.S. Department of Energy. The United States Government retains and the publisher, by accepting the article for publication, 20 acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan).

1) How do different decomposition base rates in two different decomposition models affect litter turnover times as compared to LIDET observations?
2) how are these turnover times affected by seasonal patterns in soil conditions and nitrogen limitation?

Model description and simulation procedure
Here we use the Energy Exascale Earth System model (E3SM) land model version 0 (ELMv0), which is based on the Community Land Model version 4.5 (Oleson et al., 2013). ELMv0, like CLM4.5, includes vertically resolved 140 soil biogeochemistry over 10 soil layers, the explicit representation of ammonium and nitrate pools, and associated nitrification and denitrification flux of N (Koven et al., 2013). CTC and CNT, the two decomposition submodels analyzed in this study, each have three litter pools that represent labile, cellulose, and lignin litter fractions. New litter is allocated to these pools based on type of litter (live or dead wood, fine root, and leaf) and the plant functional type (PFT). Each litter pool turns over to a separate soil organic matter (SOM) pool with a base rate that is modified 145 by temperature, moisture, oxygen availability, depth and competition for available nitrogen among plants and other microbial immobilization pathways. There are three SOM pools in CNT and four in CTC, which decompose either to other SOM pools or the atmosphere with base rates that are modified by temperature, moisture, oxygen availability and depth. For each pool transition, there is a loss of carbon in the form of CO2 flux to the atmosphere, defined as a respiration fraction. CTC parameters are based on measured values from a series of mesocosm 150 experiments (Thornton and Rosenbloom, 2005), while CNT parameters derive from the CENTURY model (Parton et al., 1988). Parameters for both decomposition models are given in Table 3.  Competition for nitrogen between plants and microbes is resolved at each half-hourly model timestep using a relative demand approach (Thornton et al., 2007). Plant nitrogen demand is driven by the portion of carbon uptake that is allocated to structural pools and associated stoichiometry given fixed carbon to nitrogen ratios for those differences in carbon to nitrogen ratios between these pools and the respiration fraction. The potential litter decomposition, which drives the microbial nitrogen demand, is a function of litter chemistry, pool size, soil moisture and temperature. If the total demand for nitrogen is greater than the available soil mineral nitrogen, plant and microbial uptake are then scaled down by multiplying their fractions of total demand by the total available 165 mineral nitrogen. In this way, both plant gross primary productivity (GPP) and litter decomposition are downregulated by nitrogen limitation. Effectively, nitrogen limitation serves to modify the base litter turnover rates in the decomposition model along with temperature and moisture conditions. Here we use the capability of ELMv0 to run with both the CTC and CENTURY-based (CNT) decomposition submodels with vertically resolved soil biogeochemistry described in Koven et al. (2013), referred to here as ELM-CTC and ELM-CNT respectively.

170
The vertically resolved ELM-CTC configuration is also the default soil biogeochemistry parameterization for ELMv0 and later versions.
For this study, we perform ELM-CTC and ELM-CNT simulations to calculate the environmental conditions in terms of the temperature and moisture scalars, as well as the background nutrient limitation conditions at each of the LIDET experiment sites. The simulated LIDET experiment at each site then takes place in a model "functional 175 unit", which is a standalone representation of the decomposition model (section 2.3). In ELM, each of the 18 LIDET sites is represented as a single gridcell in a multi-site ensemble simulation, similar to the model setup in ELMv0 in Ricciuto et al. (2018). Both ELM-CTC and ELM-CNT were run to steady state using pre-industrial carbon dioxide concentrations (constant year 1850), nitrogen deposition (constant year 1850) and climate forcing (cycling the years 1901-1920) from the Global Soil Wetness Project 3 (GSWP3; Dirmeyer et al., 2006). The 180 spinup procedure follows the accelerated decomposition and regular spinup techniques described by Thornton and Rosenbloom (2005) GSWP3 climate forcing is adjusted such that its long-term mean annual temperature and precipitation matches the reported site average given by Adair et al. (2008). ELM plant functional types are set to site-reported values following Bonan et al. (2013). Land-use change is not considered in these simulations. The LIDET experiment covers the period from October 1990 through the year 2000. During this time, soil moisture and temperature scalars (Wscalar and Tscalar , respectively), and the fraction of potential immobilization (FPI) as simulated by ELM-CTC and rates for litter decomposition (defined at 25C and ideal moisture conditions), and serve as boundary condition inputs for the decomposition functional unit.

Functional units and simulation procedure
In ELM, it is not possible to represent the small scale of the litter bags (10x10cm) in a simulation that includes a 195 big-leaf representation of vegetation at canopy scales. One would have to assume the litter bag is covering a much larger spatial area, which may cause unrealistic feedbacks to vegetation (e.g. changes in plant nitrogen uptake due to the influence of the litter bag). However, the decomposition submodel is not subject to canopy-scale assumptions. Therefore, a functional representation of the ELM-CTC and ELM-CNT decomposition models is developed. In this framework, the decomposition can proceed in the simulated litter bags using the environmental A true functional testing platform allows for the testing of specific submodels using existing ELM output and the original model subroutine code, allowing new insights about the behaviour of these submodules (Wang et al., 2015). However, due to large memory and communication requirements, this platform is not computationally 210 efficient to run over large combinations of sites and parameter values. Therefore, we developed a python-based version of the decomposition model, which is tested against the decomposition functional unit that uses the original ELM code for fidelity (Yao et al., 2019). The CTC and CNT versions of the functional unit submodel are referred to as CTCf and CNTf respectively. The LIDET model experiments are then performed using these functional units with the site-level ELM simulations providing the soil environmental and nutrient conditions (Wscalar, Tscalar and FPI 215 from the top soil layer) as inputs ( Figure 1). For each site and each leaf litter type, CTCf and CNTf are initialized with the leaf litter added to litter bags at the beginning of the experiment, divided into the 3 litter pools (labile, cellulose and lignin) according to the given litter chemistry (Table 2). We assume the temperature and moisture conditions are the same in the litter bag as in the first soil layer in the canopy-scale ELM simulation. As the simulations progress over a 10-year period, the litter decomposes into SOM, which also decomposes, releasing That full chain of decomposition is assumed to occur within the litter bag. We allow for the possibility that mineral nitrogen can enter the functional unit system (the bag) over time, and this flux is assumed to originate from the adjacent native litter and soil system. There is likely to be a local influence of processes in the litter bag on the available nitrogen and nitrogen limitation to decomposition. Therefore, we also calculate an internal value of FPI, 225 referred to as FPIlocal. Plant demand is assumed to be zero in the litter bag, and FPIlocal is calculated in the following way at each model timestep: Where litlocal is the density of remaining litter (g m -3 ) in the litterbag as calculated by the model functional unit and assumed to be in the first soil layer, and litsite is the density of litter (g m -3 ) in the ELM site simulation in the first soil layer.
A total of 108 simulations (18 sites * 6 litter types) are performed for each functional unit. Daily output is saved 245 from the CTCf and CNTf over the course of the 10-year simulated experiment, including all litter and SOM carbon and nitrogen pools. These outputs are aggregated to produce the fraction of initial carbon mass remaining, and the fraction of original nitrogen present for comparison against observations.

Sensitivity analysis
We also use CTCf to investigate the impacts of parameter uncertainty in the decomposition model using a global

Comparison between CTCf and CNTf default models
When using default model parameters (table 3), both the CTCf and CNTf models generally reproduce the observed temporal patterns of carbon decomposition averaged over the six litter types (Figure 2). The CTCf model displays 285 closer agreement with the observations for the tropical, conifer and deciduous PFT sites while the CNTf model agrees more closely for the tundra and boreal forest PFTs. This is reflected in the lower root mean squared error (RMSE) and bias values averaged over all litter types for CTCf or CNTf for those corresponding PFTs (table 5).
For the humid grasslands, although CNTf performs somewhat better than CTCf, both models decompose the litter too quickly compared to observations and are therefore biased low in their predicted carbon remaining by at least 290 10%. For all PFTs, CTCf shows faster decomposition rates compared to CNTf, reflecting the higher base decomposition rates in CTCf for labile, cellulose and soil organic matter pools. CTCf has a slightly lower base decomposition rate for lignin. Base decomposition rates occur under ideal moisture conditions at 25 o C, and without nitrogen limitation.   The actual decomposition rates are heavily influenced by the nutrient and environmental scalars ( Figure 3).

305
Environmental scalars representing temperature and moisture effects on decomposition are similar between the two models, although slightly higher in CTCf for all PFTs. Average annual scalar values range between 0.1 and 0.75 in CTCf, reflecting over a sevenfold difference in litter and SOM turnover rates depending on soil conditions (Table   6). A similar range of scalar values is observed for CNTf. The slightly lower values in CNTf reflect the use of an arctangent temperature function in CNTf rather than the Q10 function that is used in CTCf. For all PFTs, the nitrogen 310 scalar is lower in CTCf than in CNTf, indicating higher nutrient limitation in CTCf for litter decomposition. In ELM, plants and microbes compete for soil mineral nitrogen using a relative demand approach at every model timestep. Potential litter decomposition rates are influenced by the environmental conditions but do not reflect nitrogen limitation. Net primary productivity is similar between the two models, with differences within 5% at 14 sites and within 10% at the remaining 4 sites (Table 6). Therefore, litter inputs and plant demand are similar and 315 the higher nitrogen limitation factor in CTCf results primarily from higher potential immobilization caused by the higher base decomposition rates. This higher nitrogen limitation thus reduces the effective difference in actual decomposition rates between the two models.   The percentage of the original litter nitrogen is plotted as a function of the percentage of original carbon remaining ( Figure 4). While the percentage of carbon remaining is always declining over time, the percentage of nitrogen remaining may increase due to immobilization of soil mineral nitrogen from external sources outside of the litter 335 bag. This is most evident in the litter types with high carbon to nitrogen ratios, for example PIRE and THPL.
These types incur higher immobilization demand when decomposing from litter to SOM. In the functional unit representation, the litter bag is assumed to have no influence on its surrounding environment (i.e., it does not impact plant or immobilization demand outside of the bag). Therefore, immobilization demand is assumed to be met by external sources of nitrogen but is limited according to the calculated value of FPI from ELM and locally within 340 the bag (equation 2). For both CTCf and CNTf, the models strongly overpredict the amount of nitrogen present in the litter bags for the PIRE, THPL, ACSA, and QUPR litter types. For the DRGL type, the models better predict the behaviour because the carbon to nitrogen ratio is low and there is little immobilization. Despite the different litter and SOM decomposition base rates, the models behave similarly because they have similar carbon to nitrogen ratios for SOM and similar respiration fractions (Table 3), meaning that the total immobilization demand integrated

Model sensitivity analysis
We conducted a global sensitivity analysis of two model output variables to 11 CTCf parameters (Table 4). For the fraction of carbon remaining, the sensitivities are plotted for each of the 10 years of the experiment ( Figure 5).
In the stacked bar plots, the height of each bar represents the main effect sensitivity index for a particular parameter.
perturbations of that parameter result in positive or negative effects on the output of interest (i.e, positive or negative correlations). For the tropical site BCI (Figure 5a), the fraction of carbon remaining is sensitive in year 1 to five parameters: the fraction of potential immobilization outside the litter bag (FPI_outside), the respiration fraction multiplier (rf_mult), the multiplier for SOM decomposition rates in pools 1-3 (k_som123_mult), the multiplier for  Over the 10 years of the simulations, significant changes occur in the parameter sensitivities. At BCI, we see that the relative importance of FPI_outside decreases quickly and is no longer sensitive after three years. At this stage, a large fraction of the litter has already transitioned to SOM. Therefore, there is less immobilization demand, and additionally some of that demand may now be met by mineralization occurring within the bag. Also because of litter transitioning to SOM, k_lit_mult becomes less sensitive while k_som123_mult become more sensitive. 375 rf_mult becomes increasingly important during the first three years and remains sensitive through the simulation period. The respiration fraction parameter continues to be important as it applies not only to litter to SOM transition, but cascading SOM transitions as well (e.g., SOM1 to SOM2 and SOM2 to SOM3). Towards the end of the simulation, especially in the last year, the carbon remaining becomes marginally sensitive to k_som4_mult, which is the longest-lived SOM pool with a base turnover rate of 27 years. Large differences in parameter sensitivities are also seen across sites. The contrast in sensitivities between BCI ( Figure 5a) and BNZ (Figure 5b) reflects the extreme difference in environmental conditions between the tropical and boreal sites. FPI_outside is much more sensitive at BNZ and remains highly sensitive for a longer period of 390 time. Therefore, litter decomposition in boreal systems may be more sensitive to the conditions in the surrounding ecosystem. There is also increased initial sensitivity to k_lit_mult, and a longer time to transition to higher sensitivity of k_som_mult, reflecting the slower decomposition rates at BNZ. The fraction of litter that is labile (litter_flab) is also more sensitive initially at BNZ than at BCI. The default base rate for the labile litter pool is less than one day, so that at BCI after one year there is likely very little labile litter left. However, at much colder 395 and drier BNZ, some labile material may remain especially in simulations with low values of FPI_outside (as low as 0.03) and low k_lit_mult (as low as 0.2) that could combine to produce an effective decomposition rate 150 times slower than the default parameterization with no nutrient limitation. While the magnitudes vary, the temporal patterns of sensitivity and correlations with the carbon remaining are similar between the two sites. For the deciduous site HFR (Figure 5c), the sensitivities generally fall in between BNZ and BCI with similar correlations 400 and temporal patterns.
We also calculate the sensitivity of the fraction of nitrogen remaining as a function of time for the same three sites ( Figure 6). This quantity of interest is the most sensitive to the litter carbon to nitrogen ratio (litter_CN) for all sites and years. This parameter is the most important in determining how much immobilization occurs over the simulation. Higher carbon to nitrogen ratios lead to increased immobilization, and greater amounts of nitrogen 405 remaining. This high sensitivity is also reflected in the large differences across litter types seen in the default model  In addition to the sensitivity analysis, we also plot the range of predictions of the CTCf ensemble at the same three 420 sites for three litter types (Figure 7). In these simulations, the litter chemistry parameters were held constant at the values specified for the litter type of interest (table 2) while the other eight parameters were allowed to vary over their ranges (table 4). The purpose of these simulations is to determine the relative contributions of parameter and model structural uncertainty to the predictions. The combinations of parameters produce a wide range in the predictions of the fraction of carbon remaining. The Pinus resinosa (PIRE) type, which has the highest C to N 425 ratio of the three types shown, has the largest spread at all three types. The predictions for the default CTCf and CNTf models are also shown for comparison. For all cases except one, the uncertainty range of parameters in CTCf leads to a range of predictions that encompasses the CNTf default. For the Acer saccharum (ACSA) at the deciduous site, CNTf retains more carbon than CTCf for any combination of CTCf parameters for the first four years of the simulation. This may be due in part to the relatively high labile fraction of ACSA compared to the other litter types 430 (table 2). Because of the very fast turnover of labile litter, and the fast cascade through the SOM pools with carbon lost at each transition, this carbon is lost from the system quickly with any combination of the uncertain parameters.
However, in CNTf, in addition to the slower turnover, there is a more complex set of transitions between SOM pools (Koven et al, 2013). Some of the matter from the first soil organic matter (SOM1) pool transitions directly to the slowest pool (SOM3) rather than cascading to SOM2, meaning that a larger fraction of the originally labile 435 pool may persist. Similar behavior is seen at the boreal site BNZ; the range of uncertainty in CTCf only barely includes the CNTf prediction.  Table 4, with the litter chemistry parameters held constant for the specific litter types. Results are plotted for three litter types and three sites: BCI (first row), BNZ (second row), and HFR (third row). For the percentage of original nitrogen remaining as a function of carbon remaining, the range of uncertainty in CTCf always includes the CNTf prediction (Figure 8). The predictions for the PIRE and ACSA litter types have 445 large ranges of uncertainty for all sites: For these two types, the peak amount of nitrogen is between 300% and 400% of the original in the highest cases while in the lowest cases it is barely over 100%. This large range reflects a high uncertainty in the total amount of immobilization, even when the litter chemistry parameters are specified for specific litter types. Therefore, the rf_mult, cn_som and k_som_mult parameters are largely responsible ( Figure   6). However, a substantial number of observations include values below 100% at the same levels of carbon 450 remaining ( Figure 4). This may indicate structural errors in CTCf because even with the large range of parameter uncertainty, the model predictions do not encompass the observations. A much smaller range of uncertainty is seen for the Drypetes glauca (DRGL) type, which has a carbon to nitrogen ratio of 24.2. For most combinations of parameters, there is little or no immobilization of nitrogen required for this litter type. Parameter ranges are defined in Table 4, with the litter chemistry parameters held constant for the specific litter types. Results are plotted for three litter types and three sites: BCI (first row), BNZ (second row), and HFR (third row).

Discussion
resolved soil organic matter and litter decomposition (Koven et al., 2013), which was a subsequent model 490 development not tested in the Bonan et al. (2013) analysis. Due to the effects of higher inputs of litter into the top soil layers combined with the slow process of vertical diffusion that moves SOM to deeper layers, immobilization demand tends to be much higher than mineralization in these top layers. Therefore, values of FPI are much lower in the first soil layer (that is used to provide the boundary condition for the LIDET experiment) than for a columnlevel average. It is also important to note that although representing soil carbon with a single layer led to low biases 495 in global SOC with CTC in the CESM framework (Todd-Brown et al., 2013), using CTC in ELM and E3SM with vertically resolved SOC produces global SOC stocks more in line with observations (Burrows et al., accepted).
Another discrepancy between our analysis and that of Bonan et al. (2013) is that our analysis of CNTf differs in a few key ways from DAYCENT. First, CNTf differs from DAYCENT because it was adapted to better match the existing model structure in CLM4.5 and ELM. Although the litter and SOM pool structures, flows and turnover 500 times in CNTf closely match DAYCENT, the carbon to nitrogen ratios for SOM are fixed parameters in CNTf while in DAYCENT these ratios increase with low soil mineral nitrogen. The specified values for these carbon to nitrogen ratios are on the low end of the ranges used in DAYCENT (Parton et al., 1998), and are similar to those used in CTCf (Table 3). Although we did not test the parameter sensitivities of CNTf, we note that the carbon to nitrogen ratio and the respiration fraction are sensitive parameters for the fraction of nitrogen remaining in CTCf. This may 505 explain why CTCf and CNTf behave similarly for this variable (Figure 4), although DAYCENT performed significantly better than CTCf in Bonan et al (2013). Adjusting these parameters or allowing them to be variable in time may improve the performance of both models. Second, setting the FPI boundary conditions for the functional units with ELM simulations also causes important differences for the comparison between the two models. Because of the difference in base turnover rates, immobilization demand is generally lower in CNTf than 510 CTCf, leading to higher FPI values in CNTf and effectively reducing the difference in actual litter turnover rates between the two models. This suggests that nitrogen availability is a limiting factor on litter decomposition, which is consistent with experimental studies (Cleveland et al., 2006). Therefore, under certain conditions, the model turnover rates and structure for litter decomposition is less important than how nitrogen mineralization and external nitrogen inputs are represented. Finally, we attempt to more closely represent the unique environment within the 515 litter bag. Inside the bag at the beginning of the experiment, there is likely to be higher unmet immobilization demand because of the relatively high C:N ratios and lack of SOM. Therefore, we believe the weighted FPI (equation 2) is a more realistic representation of the nitrogen limitation inside the litter bag. Generally, this value of FPI is lower than the external FPI at the beginning of the experiment and contributes to slower decomposition.
A key parameter sensitivity in CTCf is the respiration fraction multiplier (rf_mult). This parameter sets the carbon 520 loss for the transitions between litter and SOM and between different SOM pools. Therefore, for the LIDET simulation experiment it is an important control on the carbon remaining in the system over time and on the amount of immobilization demand caused by litter decomposition. These respiration fraction parameters are closely linked (inversely) to carbon use efficiency (CUE), which describes the metabolic efficiency of microbes in the decomposition process. While the respiration fractions for specific transitions are constant in CTCf and CNTf, 525 there is strong evidence for the dependence of CUE on environmental and stoichiometric parameters (Manzoni et al., 2012;Sinsabaugh et al., 2013). Some of this dependence can be captured by the variation in respiration fractions among different pool transitions, but the model is limited in its capability to represent CUE dynamics because it does not explicitly represent microbial biomass or communities. In the absence of these mechanisms, we may consider CUE (and therefore respiration fraction) as uncertain parameters to be best fit for the observations of 530 interest. However, CUE may be positively or negatively impacted by different aspects of climate change (Manzoni et al., 2012), so that predictions of litter and soil carbon dynamics under climate change scenarios using fitted CUE parameters under present day conditions may be inaccurate. Including a more mechanistic representation of CUE in Earth system models has been increasingly recognized over the last several years an important goal. A number of model development efforts are underway to represent microbes explicitly in globally relevant 535 modelling frameworks that are being or may eventually be included in Earth system models. Such models may more accurately represent the dynamics of CUE and therefore also improve the representation of SOM. For example, the microbial-enzyme-mediated decomposition model (MEND) explicitly represents microbial biomass carbon and two key enzymes in the decomposition process (Wang et al., 2013;Wang et al., 2014). The MEND model includes physically measurable pools and dynamic CUE, which can change the sign of soil organic carbon 540 response compared to the assumption of constant CUE in a warming scenario (Wang et al., 2013). Globally, the temperature sensitivity of CUE may regulate the long-term response of soil carbon to warming (Li et al., 2014).
Another SOC model, Carbon, Organisms, Rhizosphere, and Protection in the soil environment (CORPSE) can simulate root-microbe interactions (e.g., priming effects) and the formation of protected SOM pools that may occur under certain conditions (Sulman et al., 2014). A global-scale comparison of CORPSE and the Microbial-545 Mineralization carbon stabilization model (MIMICS) model (Wider et al., 2014), which both represent microbes explicitly, shows disagreement on the magnitude of soil carbon stocks and their long-term fate (Wieder et al., 2018).
Oscillatory behaviour is also observed with many of these microbially explicit approaches (Wang et al., 2014b, Li et al., 2014. More model-data integration may be necessary to constrain these increasingly mechanistic models to produce consistent and credible responses over long timescales. MIMICs was also compared to selected LIDET observations with favourable results (Wider et al., 2014). Unfortunately, the breadth of measurements made in the LIDET study is not comprehensive enough to constrain these complex models, therefore, a new LIDET type study that includes measurements of microbial biomass, communities and enzymes may be of great benefit to the community.
In addition to the lack of explicit microbes in CTCf and CNTf, other structural shortcomings may contribute to 555 biases seen in the models. A large number of observations indicate the fraction of nitrogen remaining falling below 100% for intermediate values of carbon mass remaining during which the CTCf predictions are always above 100% for any combination of parameters. This may indicate a mechanism for nitrogen loss that is not currently represented in ELM. Soil mineral nitrogen may be lost through denitrification and leaching, but these terms tend to be very small in the first several years of the simulations due to high immobilization demand. Leaching of 560 organic nitrogen may be an important loss pathway from the litter bags, although it is not usually considered in terrestrial nitrogen budgets (Neff et al., 2003). Including this effect in the model may improve predictions. The "home field" advantage effect may also cause biases in model predictions. This effect refers to litter types that are introduced into ecosystem where they are native decomposing faster because the microbial communities are present that prefer these types of inputs, whereas non-native litter types may be less preferred. This was found to be an 565 important factor for DRGL in the LIDET experiment (Gholz et al., 2000). Parton et al. (2007) also noted that the relationship between CDI and decomposition rates did not work well for arid grasslands as for other biomes in the LIDET study and hypothesized that ultraviolet radiation may accelerate decomposition in these systems. This is supported by other studies in arid systems (e.g. Gallo et al., 2006, Brandt et al., 2007. For this reason, we did not include the arid grassland biome in our study. However, UV radiation may play a role in other ecosystems as well, 570 including in mesic grasslands. A litter decomposition study designed to measure the role of UV photodegradation (Brandt et al., 2010) reported significant UV effects at the Cedar Creek (CDR) site, a humid grassland system also used in the LIDET study.

Conclusions
To predict carbon cycle feedbacks to climate change, it is critically important to model the decomposition of litter 575 and soil organic matter accurately. A functional unit modelling approach was designed to simulate the LIDET decomposition study using two decomposition models in the ELM framework. We found that the converging trophic cascade (CTC) decomposition model, which was parameterized with a series of mesocosm experiments, reproduces the observed patterns of decomposition in LIDET reasonably well when driven with values of soil moisture, temperature, and nutrient limitation that are internally consistent with ELM simulations. The introduction of vertically resolved SOM in CLM4.5 and later ELM helped to correct apparent biases related to previous singlelayer implementations of CTC that resulted in too rapid litter decomposition and underpredicted soil carbon stocks.
When both decomposition models are implemented in ELM and tested against LIDET observations, CTC is not substantially different from the DAYCENT-based CNT model; the spread of predictions resulting from parameter uncertainty is equal to or greater than the differences caused by model structure. Modelled outputs were highly 585 sensitive to the respiration fraction, highlighting the importance of accurately simulating carbon use efficiency. In the future, microbially explicit modelling frameworks currently under development in Earth system models would benefit from litter bag experiments coupled with observations of microbial population dynamics.