The incorporation of the Ecosystem Demography concept significantly alters the representation of the land surface in the CLM. The purpose of the changes is to represent, in a discretized manner, the disturbance-driven biotic heterogeneity. In the CLM(ED), the new tiling structure represents the disturbance history of the ecosystem. Thus, some fraction of the land surface is characterized as “recently disturbed”, some fraction has not experienced disturbance for a long time, and other areas will have intermediate disturbances. Newly disturbed areas are generated periodically and mechanistically by events such as fire or the falling of large trees. The patchwork of different stages of succession within a given geographical area is discretized into a set of similar “disturbance history class” units. Note that within each of these disturbance history classes may exist a variety of plants of different types, each of which may have different ages themselves. This formulation is described next .

Representing the heterogeneity of plants is challenging in ecosystem models operating the Earth system scale, considering the variability and myriad physiological attributes, sizes, and spatial positions of real plant populations. One way of addressing this heterogeneity is to simulate a forest of specific individuals, and to monitor their behavior through time. This is the approach taken by “gap” and individual-based models (IBMs), e.g., LPJ-GUESS , SEIB and SORTIE . Their increased computational requirements mean that these models typically use a daily time step for gas exchange calculations, while the Community Earth System Model, and most other ESMs, require gas exchange to be calculated at 30 or 60 min resolution . For the sake of computational efficiency within this framework, the ED model takes the approach of grouping this hypothetical population of plants into “cohorts”. Cohorts are discrete groups of plants, which are essentially clones of each other, and are differentiated from other cohorts primarily by their plant functional type and size. Each cohort is associated with a number of identical trees, ncoh (where coh denotes the identification or index number for a given cohort).

In each disturbance history class, the hypothetical population of plants is divided first into discrete plant types consistent with the standard approach to representing plant diversity in large-scale vegetation models. In addition to this, the ED model also groups plants into numerous size classes, thus enabling vertical interactions. Cohorts of the same functional type may co-exist and compete in the same shared space as different sizes. The exact nature of the size classes emerges from the cohort fusion routines, discussed in Supplement A. Importantly, for each plant type/size class combination, the properties of the cohort's representative individual plant are maintained and prognosed (numerically integrated through time). These properties can be thought of as an average for the group of plants represented by the cohort. Note that competition for below-ground resources, namely water, remains affected only by vertical root distribution, and is unaffected by the introduction of the ED concept into the CLM. All plants have access to the same water pool, as described in Supplement A.

Traditional DGVMs prescribe only one single average individual of each PFT without the use of the cohort concept; thus, the ED approach represents a compromise in representation of forest dynamics between these two approaches. Other “cohortized” forest models exist in the literature, notably, GAPPARD , TREEMIG , the PPA model and later versions of the LPJ-GUESS model (e.g., ), but few studies (if any) have looked into the comparative merits and drawbacks of these different approaches.

We focus here on the problem of predicting the extent of evergreen and cold deciduous strategies in temperate regions. Deciduous and evergreen trees vary most obviously in their approach to leaf production. Typically, deciduous trees produce thinner leaves with lower leaf carbon mass per unit area (Ma, gC m-2), or the inverse of specific leaf area, that only allow the plant to photosynthesize for the period of the year when these leaves are viable , whereas evergreen leaves typically have more expensive construction and persist year round. Leaf nitrogen content per unit area (Narea, g m-2) and productivity also vary with leaf thickness , and are thus related to Ma and leaf lifespan (Ll, years). These three properties, Ma, Ll and Narea, are among the best-quantified leaf traits in existing databases , and together can plausibly define alternative leaf construction strategies. Furthermore, at a global scale, trade-offs exist between these three properties, and it has been suggested that the existence of such constraints on parameter space represents a key opportunity to simplify the representation of vegetation within DGVM models . To investigate how parameter choice impacts the outcomes of the model, we use the GLOPNET leaf trait database to define Ma, Ll and Narea. Within plant functional types, which are defined here as evergreen vs. deciduous trees and needleleaf vs. broadleaf trees, there are large variations for all parameters within the database (Figs.  and ). Thus, there exists a problem of parameter choice for these three properties. One approach is to simply use either the mean properties of the data for each plant type , or a single linear fit of the relationship between the different variables. This approach, while compellingly simple, presupposes that the database represents an appropriate sample, either of the mean of the existing plants, or the relationships between the variables. This is, on account of sampling biases , quite unlikely to be the case; as such we take a different approach that retains the observed spread in the available data. In this study, we construct PFT-specific three-dimensional covariance matrices (Figs.  and ) that represent our knowledge of the direction and fidelity of the trade-offs between the three traits and thus define a set of plausible “proxy species” within each plant functional type, defined in this case by phenological habit (i.e., evergreen or cold deciduous). We consider all parts of the normally distributed covariance matrix to be equally likely (since their likelihoods are derived from the observed data). We then re-sample, from this distribution, a set of 15 parameter combinations for deciduous broadleaf (DBT) and 15 for evergreen needleleaf (ENT) trees, using a multivariate normal distribution sampling routine, the mvnrnd function in MatLab .

Narea values are substantially higher for ENT than for DBT. report the relationship between photosynthetic capacity Vc,max,25 (µmol m2 s-1) and Narea for DBT and ENT, and find that ENTs have much lower instantaneous nitrogen use efficiency than DBTs, using their coefficients. We thus calculate Vc,max,25 as Vc,max,25=33.79Narea for DBT, and Vc,max,25=20.72Narea for ENT. Without this modification, a naïve approach to scaling from Narea to Vc,max,25 would give ENTs a photosynthetic capacity 50 % higher than DBTs.

This model parameterization approach only modifies a small fraction of the total number of the parameters that are necessary within the CLM(ED) framework (Supplement A). To increase the tractability of the simulations and to constrain the changes in parameters between plant functional types, we kept all of the remaining model parameters constant. We acknowledge, and indeed emphasize, that the outcome of the simulations could be altered by modification of other parts of the model parameter space. Our aim here is not to derive the “best possible” simulation of biome boundaries, but more to investigate the consequences of parameter choice within a relatively small and well-constrained framework. Few other model parameters have the same density of observations ; thus, the scenario represented by Ma, Ll and Narea is one of the best test cases for deploying trait data to predict biome boundaries.

To explore the consequences of parameter choice for the fidelity of the predicted biome boundaries, we ran a series of ensemble simulations, each using one of the 15 parameter combinations resampled from the three-dimensional covariance matrix, as described above and in Table . To allow for direct attribution of biome boundary position to our hypothesis (e.g., that the relative carbon economy of deciduous vs. evergreen plants can explain their distributions), we assume here that there are no other differences between the properties of the ENT and DBT plant types. These simulations were run five times, using a control and four alternative structural assumptions described in Sect. 4.4.

Regional model runs were conducted for the eastern United States. We selected this region on account of the continent-scale biome boundary shifts evident between phenological habits along the north–south axis of this domain. In the eastern United States, there is a clear transition from evergreen vegetation in the north to heavily deciduous-dominated ecosystems in the mid-latitudes, then back to evergreen in the southern and subtropical regions. The problem of parameterization of plant functional type attributes within the context of structural variants is complex, therefore we intentionally focus on this limited-scope regional problem, to allow a more thorough investigation of the properties of the model. We acknowledge that historical land-use impacts affect this study area, but we both screen out heavily impacted areas from our analysis and only focus on forested ecosystems, reducing this impact substantially (see the latter section on observational constraints). Other clear shifts in phenological habit occur globally, most notably at the rainforest–savanna biome boundary , but methodologies for simulating drought-deciduous phenology are not as well understood as for cold-deciduous phenology , and the issue is complicated by interactions with modeled soil and plant hydrology . Future studies will investigate other biome boundaries and ultimately the properties of global simulations.

The model is forced with 6-hourly climate drivers derived from , re-gridded to a 0.9×1.25∘ resolution grid and run from 1972 to 2003 for the eastern USA (90–65∘ W, 25–50∘ N). Because of our prioritization of ensemble experiments to illustrate the dependence of modeled plant competition on parameter values and model structural variants, rather than to explore the consequences for the entire (soil, vegetation, atmosphere) carbon cycle, we ran the models until the vegetation distribution appeared stable. Because of the absence of a nitrogen cycle in our simulations, this period was relatively short (i.e., approximately 30 years). The carbon budget of the represented ecosystems was not necessarily in balance at this time, but there did not appear to be a trajectory affecting the ecosystem composition, the output variable of interest. Our other outputs of interest, LAI and GPP, stabilize well before this time. Each ensemble member was initialized from bare ground, seeded with equal numbers of saplings of each plant functional type (ENT and DBT).

To evaluate the model predictions, we use the AVHRR vegetation continuous fields (VCF) product , which assesses global vegetation patterns in terms of leaf type (i.e., needleleaf, broadleaf) and phenological habit (i.e., evergreen, deciduous). The fraction of vegetation in each class is determined for each 5 km cell, and the data were re-gridded to the same 0.9×1.25∘ model grid. We generate a metric of average observed evergreen fraction (Feg) for each grid cell. Furthermore, we also use the MODIS leaf area index (LAI) product to evaluate model performance across the simulated domain. Leaf area index is a property often used to benchmark plant physiology models because it is a critical determinant of both energy and carbon exchange processes, despite our imperfect ability to generate LAI products from canopy greenness indices . In this instance, our primary objective is to predict spatial variation in LAI at a regional scale. Further studies will be expanded into the use of other metrics of canopy greenness (e.g., fraction of absorbed PAR – photosynthetically active radiation), using CLM4.5(ED)'s increased fidelity representation of the canopy structure (Supplement A). Areas with heavy (> 50 %) influence of anthropogenic land-use change, as determined by the CLM surface data sets , are masked out in model–data comparisons, since the model is only relevant to the prediction of natural vegetation LAI. Since the VCF product only reports values relevant to forest vegetation cover, it is relevant to test the model predictions against areas with land-use change because the herbaceous/crop areas are already screened out. Finally, we also compare model outputs to the Fluxnet GPP product , which scales fluxes observed at eddy covariance measurements sites to a globally gridded product using climate and vegetation drivers. The Fluxnet GPP has previously been used to validate CLM GPP predictions , and while it relies on data that are sparse for some regions, errors for this latitude band are relatively low .

Numerous aspects of carbon cycle process representation are uncertain in land surface models, and, using our mechanistic modeling framework, these uncertainties can propagate into predictions of biome distribution. To address a subset of this uncertainty, we conducted parametric ensembles across a variety of structural assumptions pertaining to the allocation of carbon resources across evergreen and deciduous trees. We investigate the importance of assumptions related to model initialization, which is a notable determinant of final ecosystem state in models with strong positive feedbacks. We also investigate the depiction of leaf and fine root carbon economy, taking advantage of new studies that report better constraints on these processes than exist in the default model. The new data pertain to the correlation of leaf respiration with leaf nitrogen, the turnover of evergreen leaves, and the turnover rate of fine root matter. The default model setup, described in detail in Supplement A, is denoted as the control (CONT) simulation. The other four structural variants are described below.

Our four modifications give rise to a set of 24=16 potential structural combinations. Testing all 16 structural combinations for the 15-member parametric ensemble for the full eastern United States region is computationally prohibitive. Consequently, instead of testing all combinations, we add the structural modifications in one at a time to investigate the impact of each change in isolation. We therefore compute five ensembles of alternative structural variants, by adding the ALLOC, RESP, LL_TEMP and RL_TEMP changes sequentially. For each of the five variants, we run the model for 15 times with parameter values sampled from the space of Ma, Ll and Narea. The structural variants are labelled i, ii, iii, iv, and v, and are described in Table . We compare the model output to the observed data using five comparison metrics, maximum and mean annual LAI, maximum and mean annual GPP, and the single set of evergreen fraction data available.

We calculate the R2 and root mean square error (RMSE) of the spatial distribution of each metric. We acknowledge that there exists a choice of metrics (maximum vs. minimum vs. range, and spatial vs. temporal correspondence), but also note that subjectivity in the definition of objective functions is generic to high-dimensional model output .

Our analysis is concerned with the costs and benefits, or carbon economy, of the different leaf strategies. The cost of leaves is easily calculated as the investment (in terms of LMA), divided by the lifespan (in terms of LL), giving the cost in KgC per unit area per year of leaf. The benefits (in terms of carbon export), on the other hand, are more difficult to calculate, since they are manifested not only though leaf Narea and hence photosynthetic capacity, but also by the nonlinear interactions of photosynthetic capacity with environmental drivers (light, CO2, temperature, etc.). Thus, the detailed physiological model is required to generate estimates of benefit in terms of assimilation, and it is not possible to do these calculations as a simple offline analysis. Furthermore, the implementation inside the physiological model includes the impact of self-shading of leaves lower in the canopy, and thus the costs and benefits of these strategies are actually only properly assessed at the canopy scale. To address this point, we conducted additional model runs that use only one PFT at a time, using structural variant v. Using these analyses, we can assess the differences in productivity and leaf area index of the PFTs in isolation. This removes the direct effects of light competition and allows interrogation of how the competition and productivity elements of the model combine to generate the resulting distribution.

Figures , and illustrate the R2, relative RMSE and summary statistics for each structural variant and parameter combination. Figures , , and show the simulated evergreen fraction (Feg) as simulated by the different structural variants. Figures  and show the mean annual LAI and GPP of the last structural variant (run v), once all of the modifications have been made. GPP and LAI maps are shown for the other structural variants in Supplement B.

Figure  illustrates that, particularly for Feg and LAI, R2 varies primarily with structural variation, as illustrated by the horizontal striation. In contrast, variation in RMSE, particularly for GPP, illustrates the dominance of parametric variation, shown by the vertical striation in the GPP and LAI comparisons in Fig. . We did not combine the R2 and RMSE values directly, since calculating their relative weights would serve to reduce the clarity of the output exposed by using them both independently.

In the control simulation (Fig. ), every parameter combination produced a near-complete dominance by deciduous vegetation, irrespective of the variation in parameters that were extracted from the leaf trait database. The mean R2 of the predicted vs. observed Feg across the ensemble (0.04) illustrates this lack of predictive skill. Addition of the ALLOC modifications to initial leaf biomass (Fig. ) returns significant variation in predicted Feg. The model still predicts complete dominance of BDT for some parameter combinations, but also successful dominance of ENT at high and low latitudes for others. Nonetheless, only three of the simulations have evergreen cover over 25 % (where the mean for the observations is 49.2 %). The mean (and max) R2 is 0.13 (0.34), where “max” is the highest R2 value for any of the 15 parametric combinations.

The impact of altering the leaf respiration fluxes to match the observed relationship with leaf nitrogen and plant functional type had only a slight impact on the overall RMSE and R2 statistics for the evergreen fraction predictions (maps not shown on account of their similarity to Fig. ). Making evergreen leaf lifespan a PFT-specific function of temperature (LL_TEMP) has a more profound impact on the competitive ability of the NET plants at high latitudes (Fig. ). With this structural modification, seven of the simulations have evergreen cover over 25 %, and the mean (and max) R2 increases to 0.20 (0.34).

The last modification, directly including the PFT-specific impacts of temperature on fine root turnover, further increases the dominance of evergreen trees in northern latitudes, again slightly increasing the correlation with the observations. Now nine of the simulations have evergreen cover > 25 % and the mean (and max) R2 is 0.23 (0.35) (Fig. ). In general, it is clear that all versions of the model considered here display something of a systematic bias towards the prevalence of deciduous trees using this parameter space.

The impact on RMSE of the sequence of structural modifications also showed a tendency towards improvement as the average RMSE of the predicted vs. observed fraction of evergreen trees dropped from 0.48 (model run i) through 0.41 (ii), 0.41 (iii), 0.37 (iv) and 0.35 (v) (Fig. ).

The alteration of both model structure and parameters also had a major impact on the predicted LAI. This is expected, since all of the modifications and parameters are concerned with carbon economy, and realized leaf area in the model is predicted from the vertical location of the lowest leaf layer in positive annual carbon balance (Supplement A). The increase in model–data spatial coherence (R2) through the structural ensemble (from runs i to v) for Feg (see Sect. 5.2) is not echoed by changes in the R2 of mean annual LAI, which instead decreases through the ensemble from 0.45 (run i) through 0.31 (ii), 0.30 (iii), 0.14 (iv) and 0.05 (v). This trend was not apparent for the R2 of maximum annual LAI (which varies through 0.42 (i), 0.15 (ii), 0.32 (iii) 0.39 (iv) to 0.38 (v)) (Fig. ). The model error (RMSE) was also relatively insensitive to changes in the model structure, aside from the change from run i to run ii, which improved the simulations (Fig. ).

The direction of change of the R2 and RMSE statistics was not consistent due to spatial complexities. Specifically, the control simulation (run i) systematically underestimated LAI across the entire domain (Supplement B: Fig. 1) and thus had a high RMSE. The lack of much spatial structure in LAI prediction across the geographical domain, however, meant that it had a relatively good spatial coherence with the LAI data product, which is also relatively homogenous across the domain. Increasing allocation to leaf biomass in simulation ii, and thus increasing LAI overall, intensified the spatial heterogeneity of the predictions (Supplement B: Fig. 2) and thus worsened the R2, but reduced the model error.

Changing the respiratory fluxes in run iii improved the R2 fit to maximum LAI (from 0.15 to 0.32, Fig. ), potentially on account of the higher respiration rates at low latitudes acting to even out the spatial distribution of LAI (Supplement B: Fig. 3), and in doing so compensated for the decline caused by the previous modification (illustrating the possibilities of model equifinality).

Altering the leaf turnover time caused an increase in the mean LAI (from 2.66 to 3.06) by reducing canopy replacement costs at high latitudes. The model predictions thus now approach and in some cases overshoot the values observed for high-latitude evergreen forests (3.5–4.5 m2 m-2) in the data product (Supplement B: Fig. 4). In the simulations where evergreen trees are dominant, it is notable that their LAI values may be somewhat over-predicted. The final simulation (v, with the RL_TEMP modification) intensifies the reduction in tissue turnover demand at high latitudes, and thus the changes primarily amplify those imposed on LAI by the LL_TEMP modification. The model now illustrates a very wide range of potential LAI predictions, dependent on the parameters chosen to represent the ENT and DBT strategies (Fig. ). The major systematic bias in the final LAI predictions is the underestimation in the mid-latitudes of the domain. The fact that this feature is persistent across the parameter space sampled (even though there is clearly room for more detailed parameter optimization) indicates a persistent structural bias, particularly in the performance of deciduous broadleaf trees in their higher ranges. This underestimate is not substantially changed by any of the structural modifications we deploy here (all of the simulations indicate the same issue) and does not appear to result from underestimates of productivity (Fig. ), potentially implying a deficiency in carbon allocation.

It is worth noting that the LAI values predicted by the CLM4.5(ED) algorithm (which assumes leaf area optimized for net canopy carbon gain) all appear to be in the range bracketed by the observations. Historically, the CLM4.0 and CLM4.5 models have suffered from issues related to the chronic overestimation of LAI . We suggest that limiting the production of leaf layers in negative carbon might ameliorate this issue.

The correlation coefficients for GPP are consistently higher than those for LAI or for biome boundary prediction, illustrating that simulations of GPP appear generally more robust than either those for plant carbon allocation or for biome boundary prediction (Supplement B: Figs. 5 and 6). The spatial correlations of maximum annual GPP flux are relatively insensitive to the effects of structural variation (R2 values are 0.48 (i), 0.49 (ii), 0.49 (iii), 0.44 (iv) and 0.44 (v) (Fig. )). The R2 values for mean annual GPP flux are more sensitive to model structure (0.63 (i), 0.58 (ii), 0.58 (iii), 0.44 (iv) and 0.39 (v)) and, in common with the LAI predictions, decline through the ensemble.

Notably, the overall mean and RMSE values for GPP are much more sensitive to variations in parameter values than to changes in model structure (Figs.  and ), reflecting the impact of the parametric variation on the overall productivity, both directly via the impact of Narea on Vc,max, and indirectly via impacts of Ll and Ma on leaf area index.

GPP predictions using parameter setting no. 13 have a notably low R2 for mean and maximum GPP (which is actually negative for runs ii though v, resulting from the residual sum of squares being larger than the total). This simulation has the highest fractions of evergreen vegetation, and generates very high LAI and thus high GPP values in the far north of the domain (Supplement B: Figs. 7 and 8). As a result, in the latter parts of the structural ensemble, no. 13 has a notably poor spatial correspondence to the observations (which show a decline in GPP with latitude). Several of the other high evergreen cover ensemble members (nos. 5, 12, 15), all of which have an unrealistically high LAI in the northern areas, also show a degraded correspondence to the GPP data product. Not all parameter combinations show this, suggesting that some of the Ma and Narea combinations might be inappropriate for use in the far north (see discussion).

Figure  illustrates the absolute difference between the productivity (annual NPP) of the EBT and the ENT for the third year of the simulation for structural variant v. Each PFT was run in isolation to calculate these differences. Here it is clear that at the mid-latitudes, the EBTs have a significant productivity advantage, which broadly maps onto the eventual distribution of these PFTs in the competitive simulations discussed above. At higher and lower latitudes, the ENT and BDT have approximately equal productivity. Parameter choice affects the distributions of the areas where EBT has an advantage, but the pattern is consistent across the ensemble, excluding parameter combination no. 13. Looking at the performance of larger trees, where the LAI is equilibrated with productivity, and effects of initialization have disappeared (Fig. ), there are either small differences or considerable productivity advantages of the ENT type (excluding ensemble member no. 14). This implies that the EBTs gain dominance early in the competitive interaction, presumably by amassing leaf area at a greater rate than the ENTs. Thus, the representation of light competition is instrumental in producing biome boundaries in this example.

At least two large biases were indicated by the structural ensemble that were not resolved by any of the tested modifications. First, the under-performance of DBTs at the northern extent of their range, and second, the over-performance of ENTs in the far north in some of the model simulations. To address the latter, find some correlation between MAT and leaf nitrogen allocation for their set of ENT species. We did not detect a relationship between MAT and Narea in the GLOPNET data; thus, this might be a topic of future investigation. It is worth noting, additionally, that the optimality criteria with which CLM(ED) predicts the leaf area index is based on the avoidance of leaves in negative carbon balance. In cases of severe nutrient limitation, this might be only an upper bound on LAI, and alternative metrics that take into account the cost of nitrogen acquisition might be more appropriate .

The CLM4.5(ED) is designed as a trait-filtering model, in that it can predict successful vegetation types from their traits via the “filter” of environmental conditions. One central premise of trait-filtering models is that “trade-off” surfaces are necessary inputs, and, implicitly, that moving along the surface means that performance increases by some metrics, but gets worse in others. The use of a proscribed trade-off surface is illustrated in the Jena Diversity (JeDi) model . Potential plants (proxy species) are selected from a seven-dimensional trade-off surface, and the environment acts as a filter on this (large) population, reducing the realized population to those proxy species that are able to reproduce under given environmental conditions. Implicit in this methodology are the assumptions that all trade-off surfaces are fixed, and that they are independent of climatic drivers.

discuss three classes of trade-offs that may be considered in vegetation models – allocation trade-offs (investment decisions in different tissues), mechanical trade-offs (intrinsic structural properties) and empirical trade-offs that must be prescribed, in lieu of understanding of their mechanistic underpinning. In our study, the three-way trait relationship between Narea, Ma and Ll is an empirical trade-off. Contrary to observations across multiple plant functional types , the within-PFT trait relationships appear weak. Specifically, large variations in leaf lifespan and in Narea are possible for the same leaf carbon investment (Ma) (Figs.  and ). If, for example, a higher Ll value is chosen for the same Ma, the cost of canopy replacement will go down, increasing plant leaf area index, productivity, and growth. There is no downside in this model framework to having longer-lived leaves. Therefore, in this case, the trait data fail to accurately define a trait trade-off. It is possible to empirically define a surface fitted to the data, and to remove the “noise” around the central tendency of the data. This approach would necessarily reduce the tendency to select plants with very high or low relative productivity, but also would, in this case, be an inaccurate reflection of the genuine spread of the data, given the lack of adherence to clear trade-off surfaces.

Higher Narea increases both photosynthetic capacity and respiration rates, so should be subject to some degree of trade-off, depending on the climatological conditions (warm nights and long winters increase the costs of high leaf N). Nonetheless, the balance of these processes appears not to produce equivalent performance across the space defined in Figs.  and . This outcome highlights two potentially problematic issues with the trait filtering approach. The first is that costs and benefits of alternative strategies might not be represented completely by simple and easily observable trade-off surfaces. The true “cost” to plants of long-lived leaves may not be a linear function of Ma. Long-lived leaves might well, for example, require investment resources in complex and energetically expensive defensive compounds, and so an alternative axis of investment and return might be functionally more appropriate. The second issue is that trade-off surfaces might not necessarily be consistent across locations . For example, differences in the environment (e.g., temperature) might increase the potential lifespan of leaves by reducing herbivory rates and damage from solar radiation.

At present, many land surface modeling efforts use a variety of approaches to predicting plant traits, inclusive of trait-filtering , direct prediction of plant traits from their environment (e.g., allocation from ) and ideas from optimization theory (e.g., stomatal conductance, vertical N allocation). Many parallel concepts exist for how to define plant traits within advanced vegetation models , but the circumstances under which it is most appropriate to use which methodology is a topic that has not been discussed widely. To move the science of vegetation modeling forward, we argue that it will become necessary to understand under what conditions empirical “short cuts” to predict traits are acceptable and necessary, and under what circumstances detailed mechanistic prediction is either possible or desirable. In the first instance, it is, of course, imperative to both further advance the collection of data on plant traits and processes where possible, and to continue investigations into plant trait databases that already exist, ideally in a context that is linked to the requirements of predictive models (e.g., ). We consider that the analysis of plant trait data to determine how both environmental conditions and plant strategies (such as the “fast–slow” axis, proposed by ) can be used to generate robust predictive models is an extremely high priority. It is worth noting also that while our study does not consider the impact of changing climate on carbon cycle processes, the alternative structural variants imply both different lag times and feedbacks to the impact of climate, via the use, or otherwise, of direct impacts of temperature on turnover processes.

Another aspect of our study highlights the importance of ensembles for the investigation of model properties. It is the default practice, in land surface modeling and climate science generally, to present results using the name of a particular model to depict an invariant set of default parameter and structural assumptions (e.g., CLM4.5, JULES1.0, ED2) and to assess the merits of only one version of a model from the hyper-dimensional set of potentially viable model predictions. Such “simple” tests of model performance against observations, however, explicitly convolute the structural, parametric and initial condition contributions to model error, and, therefore, interpretation of mismatches with data is difficult. We here argue that increased use of both structural and parametric ensembles is beneficial for the development of understanding of complex land surface modeling schemes.

In Earth system modeling more widely, the use of initial condition ensembles is increasingly considered to be critical for the evaluation of model behavior . Model inter-comparison projects, both for Earth system models and their land surface model components , are used as a means of investigating the impact of alternative model structures, although typically the high dimensionality of the inter-model differences renders it difficult to assess the causes of differences between models (but cf. ). In this study we investigate a variety of model structures within the same framework. This approach, also adopted by , , , , , and among others, enables the differences caused by individual modifications to be quantified and understood, and therefore potentially provides a more tractable approach to understanding the processes leading to prediction differences than a standard model inter-comparison experiment. Perturbation of the parameters of land surface models (referred to as “perturbed physics” ensembles) is rarely undertaken at scales larger than one grid cell (but cf. and ) on account of the high time and energy costs of global model simulations. Perturbed physics ensembles of Earth system models have been conducted but have typically focused on processes unrelated to the land component . While some objective statistical techniques have been used for single sites , inverse model calibration of DGVMs over large regions is not yet considered a computationally tractable problem. More typical is the process of ad hoc parameterization, either using values of observable parameters from the literature that may or may not be representative of globally relevant values, or the use of “tunable” parameters that might be adjusted to bring the overall model behavior closer to observations, as also discussed by and . Thus, model parameters are typically not optimized and therefore the comparison of model performance to benchmarking data is not necessarily a good test of the structural validity of the model components . Model structural performance is therefore much more commonly assessed at individual sites, where sensitivity to parameters can be investigated more comprehensively . An alternative path forward might be to present models with no default parameter values, and instead with a range of physiologically plausible parameters, thus reducing the correspondence between named model structures and a single deterministic set of outputs.

Code for this paper is available in the CESM svn repository (registration required) at the following address: https://svn-ccsm-models.cgd.ucar.edu/clm2/branch_tags/ed_v0.1.0_tags/ed_v010_21_clm4_5_1_r097.