In the LPJmL model, the different PFTs compete for space/light, water and nitrogen. In past model versions, these resources were distributed between PFTs based on foliage projective cover (FPC). The FPC is used as a proxy for actual cover, which would require the explicit simulation of the plant geometries. Distributing these different resources based on one variable neglected the importance of different traits for the uptake of different resources. In particular, water uptake should also be dependent on root traits such as the extent of the root network and the amount of fine root biomass . Using root traits to determine access to water enables the model to simulate different strategies for water-resource use. Therefore, we adapted the implementation of water supply to make it dependent on root biomass instead of FPC to provide a distinction between the criteria for aboveground and belowground resource uptake and distribution. Based on the concept of the FPC, we implemented a belowground equivalent based on root instead of leaf biomass (Appendix ). First, the PFT's access to water from different soil layers is calculated as described in . Second, the amount of water available for the PFT is determined by considering its root biomass and the new parameter (kroot), which is a proxy for root properties associated with morphological properties of the root network (e.g. branching and spread).

The LES describes correlations between several plant functional traits. Among these are the specific leaf area (SLA) and the leaf longevity, which can be used to express the differences between resource acquisitive vs. resource conservative growth strategies . The resource acquisitive strategy is associated with fast growth of leaves at low construction costs with a high SLA and a short longevity. In contrast, the resource conservative strategy promotes slow growth of long-lived leaves with low SLA. Therefore, to represent the trade-offs associated with the differences between these strategies, a functional relationship between SLA and leaf longevity can be used.

Despite the importance of SLA and leaf longevity for several processes within LPJmL, the SLA vs. leaf longevity trade-off has not been implemented for managed grasslands in LPJmL before. SLA is used to calculate the leaf area index (LAI) of a given grassland area from the dynamically computed leaf biomass, which is important for the interception of light energy and thus for photosynthesis. The leaf longevity was represented through turnover rates, which determine the amount of leaf biomass transferred to the litter layer . As long as differences between ecological strategies were not considered and only one PFT was used to simulate a managed grassland, this approach was sufficient. However, this means that grasslands along a resource stress gradient only differed in their productivity but not in other aspects of the community. Yet in reality, slow-growing, resource-conservative plants in stress-prone ecosystems are not only less productive but also supply less forage with a lower nutrient content . Such ecosystems are also more vulnerable to overgrazing and recover more slowly from disturbances . Incorporating the SLA vs. leaf longevity trade-off is essential to account for the differences between ecological strategies, which are important to adequately represent ecosystem functions of managed grasslands under different climatic conditions and management.

The SLA vs. leaf longevity trade-off has already been implemented in the related LPJmL-FIT model and applied to tropical and European forests . For this study, we implemented the SLA vs. leaf longevity trade-off for managed grasslands using a functional relationship between the two based on trait observations. Similar to , we derived a power law for SLA and leaf longevity from trait data retrieved from the TRY database . This power law provides a functional relationship between SLA and leaf longevity, which is used to calculate the PFT-specific leaf longevity from predefined SLA values within LPJmL-CSR (Appendix ). Based on the alignment of the resource conservation axis of the root economic space and the LES , we assume that leaf and root longevity are not independent from each other and maintain a fixed ratio of the two in LPJmL-CSR.

Herbaceous plants are adapted to different growing conditions and therefore have different reproduction strategies and whole plant – or for graminoids phytomere – longevity. In LPJmL, each herbaceous PFT was simulated using only one average individual with specified properties. Age mortality was implicitly included in the representation of turnover of leaves and roots and not as a separate process. The only additional cause of mortality was negative leaf and/or root biomass after allocation as a result of prolonged stress. While this may be caused by water stress, additional causes of mortality from water stress such as embolism or heat stress were not considered.

LPJmL does not simulate seed bank formation, and reproduction is not limited by the amount of seeds available in a seed bank. Instead, the establishment depends on the bare-ground area and the PFT-specific establishment rate. Furthermore, in LPJmL 5, reproduction was simulated as a biomass increase of the average individual. We argue that this was not sufficient to simulate different reproduction strategies, which differ in the amount of seeds, seed survival and germination rates, and germination requirements .

In the representation of CSR strategies in LPJmL-CSR, we retained the approach of establishing seedlings instead of seeds but allowed PFTs to establish different numbers of seedlings in agreement with their reproductive strategy. To achieve this, we abandoned the approach of using only one average individual to simulate each PFT and introduced a dynamic number of average individuals assuming a homogeneous population (i.e. individuals of the same PFT share the same properties) but form the community together. Based on the existing implementation, we modified the reproduction so that additional individuals are established and thereby increase the number of average individuals simulated. Each day, the number of average individuals of each PFT is increased if there was bare-ground area available. The bare-ground area is distributed between established PFTs depending on their establishment rate kest. The total amount of seedlings established is calculated based on kest, accounting for the bare-ground area. Subsequently, the number of average individuals is increased and the individual-specific carbon and nitrogen stocks are adjusted. LPJmL-CSR does not consider trait plasticity or evolutionary processes and therefore does not account for phenotypic adaptation. This also means that already established and newly established average individuals share the same traits. Since space for plant establishment is limited and age-related mortality is common in natural grasslands , we prohibit an infinite increase in the number of average individuals by adding an age mortality based on the growth efficiency to reduce the number of average individuals (Appendix ). The growth efficiency is the ratio of the net change in the individual carbon stocks (the result of net photosynthesis and turnover) and the individual carbon stocks. Assuming that old plants grow more slowly, this is used as a proxy for population age and resulting age mortality. We did not implement additional causes of mortality such as drought or fire . While the new approach does not simulate individual or phytomere morphology explicitly, it provides some implicit information on community structure and plant size through the number of average individuals, the area covered by them and their biomass. It can be assumed that few individuals that maintain a high cover and biomass must be larger than more individuals that provide a similar cover and biomass.

We assumed that the position of a PFT within the CSR triangle can be determined through trade-offs between plant functional traits along the stress and the disturbance gradients according to the relations described below. Names, descriptions and usage of the model parameters are based on the model versions LPJmL 4 and 5 .

According to CSR theory, stress is defined as constrained metabolic efficiency limiting biomass production and can be caused by a variety of factors . The stress gradient expresses the intensity of stress a species is exposed to in a certain habitat. It ranges from unstressed to severely stressed and can include the combined impacts of several stressors. Different traits and their values are associated with the ability of a plant to cope with the different stress levels. The traits of the LES together with different strategies for resource use can be used to distinguish C and R strategists (low stress tolerance) from S strategists (high stress tolerance).

Since the LPJmL model only represents a subset of possible stress factors (Sect. ), only stress arising from temperature and water as well as nitrogen availability can be considered. Within LPJmL-CSR, some traits are linked to a general response to stress independent of the stressor, while others are used to represent adaptation to specific stressors. Since the grassland steppe sites simulated by us are predominantly limited by water, we decided to focus on water stress. This allows for a better understanding of the underlying processes and the resulting patterns. To represent the stress gradient, we used functional traits associated with the growth rate and water-resource use. We selected the maximum transpiration rate (Emax), the minimum canopy conductance (gmin), the specific leaf area (SLA) and the leaf-to-root mass ratio (lmro). SLA

The specific leaf area is the ratio of leaf area to leaf dry mass and a measure of the amount of biomass required to produce a unit of leaf area. It is predominantly associated with the stress gradient in the CSR theory. SLA is used in four processes of LPJmL-CSR. First, it is used to calculate the LAI, which controls light interception and thus productivity determining the area occupied by a PFT in competition with other PFTs. Second, SLA is used to determine the aboveground biomass of newly established seedlings from the seedling LAI (see explanation of LAIsapl). Third, it is used to determine the actual mortality rate (Appendix ). Fourth, it is used to calculate the leaf longevity controlling tissue turnover and litterfall (Sect. ). The SLA can be used to determine the trade-off between short-lived, acquisitive (high SLA) and long-lived, conservative (low SLA) leaves. In contrast, in LPJmL 5 it was only used in the first and second processes.

Similar to the stress gradient, the disturbance gradient ranges from undisturbed to severely disturbed. Reproductive traits and plant stature can be used to distinguish C and S strategists (low disturbance tolerance) from R strategists (high disturbance tolerance). Functional traits associated with reproduction and plant geometry can be used to represent the trade-off associated with the disturbance gradient. We selected the following functional traits involved in the direct interaction of the different PFTs: the root efficiency coefficient (kroot), the light extinction coefficient (kbeer), the establishment rate (kest) and the leaf area index of a seedling (LAIsapl). While seedling is the more intuitive term for herbaceous plants and we will use it throughout the paper, the subscript in the parameter name refers to saplings because it was adopted from the tree PFTs in the past. kbeer

The light extinction coefficient is a parameter describing the amount of light absorbed by a vegetation layer. It is predominantly associated with the CSR disturbance gradient, but since it is used in the calculation of the FPC, which also determines resource access, it is also associated with the CSR stress gradient. In LPJmL 5 and LPJmL-CSR, kbeer is used to determine FPC controlling the PFT-specific area share and its access to light. kbeer can be used as a proxy to distinguish large (high kbeer – rarely shaded by competitors and have high light absorption capacity) from small (low kbeer – potentially shaded by competitors and have high light absorption capacity only if dominant) stature plants and is essential for the competition for light and space.

In total, we used eight parameters to distinguish the PFTs and defined plausible ranges for their parameterisation so that different CSR strategies can be represented by the extended set of PFTs. The selected traits affect a variety of processes within the model and differentiate the C, S and R PFTs along the stress and disturbance gradients. We assumed all parameters to be independent from each other. While we are aware that SLA and the light extinction coefficient kbeer are correlated in reality because the transmissivity of leaves increases with SLA we have to treat them as independent because in LPJmL, the light extinction coefficient does not describe the transmissivity of a single leaf but of the entire vegetation layer. Stacking a high number of high transmissivity leaves may result in the same light extinction compared to a lower number of low transmissivity leaves. In LPJmL-CSR, a similar kbeer would be assigned for both cases because it represents the light extinction coefficient of the entire vegetation layer.

To parameterise the new PFTs, we had to assess the model performance for different parameter sets. We included several variables in the calculation of a likelihood (logLI): forage offtake or leaf biomass; SOC, and C, S, and R strategy covers (Table ). Data on forage offtake, leaf biomass and SOC were available from several field experiments conducted at the respective sites. For C, S and R PFT covers, data were only available for the hot steppe and we defined values based on our knowledge of the site-specific conditions to agree with CSR theory for the other sites.

We optimised the logLI using a Markov chain Monte Carlo (MCMC) method with a Metropolis algorithm . This method evaluates the performance of a sequence of sampled parameter sets. In the following, we refer to a sequence as a chain and to an iteration as a link. At the beginning of the chain, a first parameter set is drawn from a multivariate Gaussian distribution with its modes at the centre of the parameter ranges for each parameter and its variances as a fraction of the parameter ranges. A fraction of the ranges is used to limit the difference between parameter sets of subsequent links which improves the performance of the algorithm. The width of this fraction is controlled through a tuning parameter and is fixed for the entire chain, while the modes of the Gaussian distribution are updated throughout the chain if the model performance calculated as the total likelihood (logLI, Eq. ) improves. 1logLIi=logPriori+logLinki The total likelihood logLI is calculated for each link i. It consists of a prior likelihood (logPriori, Eq. ) and the likelihood of the current link (logLinki, Eq. ). 2logPriori=∑jB(θi,j,p,q) The prior likelihood, logPrior, is calculated from the prior distribution, which represents an initial guess on the resulting posterior distribution. We chose a geometrical prior distribution (B) with the shape parameters p=1+4⋅(θ^centre-θ^min/θ^max-θ^min) and q=6-p. Here, θi,j represents the parameter values of each parameter j of the current link i, θ^centre represents the values at the centre of the parameter space, and θ^min and θ^max are the lower and upper limits of the parameter space, respectively. 3logLinki=∑k-0.5⋅ysim,i,k-yobs,kσobs,k2-0.5⋅log⁡(2π)-log⁡(σobs,k) The likelihood of the current link, logLinki, is a measure of the model performance of a simulation using θi,j. The logLinki incorporates the difference between simulation results (ysim) and observations (yobs) for all variables k, also including the uncertainty of the observations (σobs). The overall likelihood (logLIi) is compared to the highest likelihood that was achieved so far (logLI) to decide about the acceptance of the current parameter set. If the difference between the current likelihood and the highest likelihood (ΔlogLI = logLImax-logLIi) is positive, the parameter set is always accepted. For negative ΔlogLI, it is only accepted if it exceeds the natural logarithm of a random number between 0 and 1. This mechanism prohibits the outcome that the algorithm is trapped in local optima. At the end of the chain, the algorithm returns a posterior parameter distributions whose modes are the parameter values with the best model performance.

We used the same parameter space for all three new PFTs but ensured that we select parameter values consistent with the traits associated with CSR theory. We prescribed a hierarchy based on our expertise (Table ) for each parameter that defines whether a PFT has to obtain a higher or lower value compared to the other PFTs. For example, S PFTs must have a lower SLA than C and R PFTs, because the S strategy is associated with slower growth and longer-living tissue than the C and R strategies.

Our approach included two steps represented by two subsequent chains. The first chain was short and used a large tuning parameter so that the sampling covered the entire parameter space and an area of good model performance could be identified. The second chain started in the area discovered by the first chain, was longer and used a smaller tuning parameter to find the optimal parameter values within the area.

We evaluated the new PFTs using the mean square error (MSE) and its components (see Appendix ). For the evaluation, we either used a different data set or split the data into different sets for parameterisation and evaluation if the number of replicates was at least eight for the majority of observations (Table ). Observations with less than eight replicates were only used for the parameterisation. For the hot steppe, we used the difference in SOC between the ungrazed and grazed scenario for the evaluation because the current representation of the processes listed in Sect.  made it impossible to simulate the overall SOC level adequately. For the cold steppe, SOC data were only available for 1 year and the common management for the examined region . While this is comparable to our extensive grazing intensity, for the intensive grazing intensity we assumed a 25 % lower SOC level. We based this assumption on who reported around 25 % lower SOC content of the topsoil under heavy grazing compared to areas without or with periods of moderate grazing.

Forage offtake of the temperate grassland for the unfertilised scenario was strongly underestimated by LPJmL 5 (Fig. S2a), and the MSE improved from 431.7 to 112.2 (Mg DM ha-1)2 in LPJmL-CSR (Fig. a). For the fertilised scenario, LPJmL 5 underestimated forage offtake less severely (Fig. S2b), and the MSE was similar with 96.4 (Mg DM ha-1)2 in LPJmL 5 to 105.3 (Mg DM ha-1)2 in LPJmL-CSR. For the unfertilised scenario, the representation of SOC improved as well. For the unfertilised scenario, LPJmL 5 strongly underestimated SOC stocks (Fig. S3a), and the MSE was reduced from 1262 to 21.4 (Mg C ha-1)2. However, it remained similar with 11 and 37.9 (Mg C ha-1)2 for the fertilised scenario (Figs. b and S3b).

Simulation results for the hot steppe presented a mixed picture showing lower MSEs for leaf biomass but higher MSEs for SOC in LPJmL-CSR compared to LPJmL 5. For the ungrazed (C0) scenario, the MSE of leaf biomass improved from 10154.1 to 1.9 (Mg DM ha-1)2 (Fig. d). Similarly, for the grazed (C1) scenario, the MSE of leaf biomass improved from 9522.5 to 40.1 (Mg DM ha-1)2. The MSE for the difference in SOC between the ungrazed and grazed scenario was lower in LPJmL 5 and increased from 6.3 to 251.2 (Mg C ha-1)2 (Fig. e). LPJmL 5 already simulated the SOC difference between the scenarios well impeding improvements through LPJmL-CSR. Furthermore, improvements in leaf biomass outweighed degradation in SOC stocks and LPJmL-CSR fits the observations better overall. However, compared to observations, LPJmL 5 severely overestimated leaf biomass and LPJmL-CSR underestimated leaf biomass (Fig. S4) and both model versions overestimated SOC in the ungrazed and grazed scenario (Fig. S5).

For the cold steppe, animal feed demand was met in both model versions for the low grazing intensity (S1). Still, the MSE for forage offtake improved from 9.5 to 8.4 (Mg DM ha-1)2 (Fig. g). For the high grazing intensity (S6), the feed demand was not always met in both models versions. Here, the MSE improved from 4.5 in LPJmL 5 to 3.8 (Mg DM ha-1)2. Both LPJmL 5 and CSR underestimated observed forage offtake for both grazing intensities but the dynamics at the high grazing intensity were captured better by LPJmL-CSR (Fig. S6). Unfortunately, replicates for forage offtake were not sufficient to split the data and no additional data were available for evaluation. Similarly, only data on SOC for 1 year, which did not distinguish between areas of different grazing intensity, were available. Since these data were already used for the parameterisation, we were not able to properly evaluate SOC. While LPJmL 5 strongly underestimated SOC for the low grazing intensity (S1), LPJmL-CSR captured the observations better but still underestimated observations (Fig. S7). Values were within the standard deviation of the observations for the low grazing intensity. For the high grazing intensity (S6), we assumed that 75 % of the observed SOC to be an appropriate estimate for calibration (Sect. ). However, both LPJmL 5 and LPJmL-CSR overestimate this reduced calibration estimate (Sect. ). The MSE was reduced from 2157.5 to 60.5 (Mg C ha-1)2 for the low grazing intensity and increased from 456.7 to 2741.5 (Mg C ha-1)2 for the high grazing intensity (Fig. h).

At all sites, the different management scenarios resulted in different parameter values for the three PFTs. For the temperate grassland, the calibration selected a less resource-exploitative strategy for the C and R PFT in the fertilised scenario indicated by the lower value for the stress gradient which resulted from higher leaf longevity (lower SLA), while the S PFT's strategy remained similar (Fig. ). Additionally, all PFTs showed a lower maximum transpiration rate (Emax) and higher investments into aboveground biomass (higher lmro). For the disturbance gradient, the C and S PFTs had a higher value in the fertilised scenario. For the C and S PFTs, this indicated that the calibration selected a strategy with less offspring (lower kest) and a more efficient root network (higher kroot). The R PFT had a lower value caused by an increase in number of offspring (higher kest).

For the hot steppe, the S and R PFTs showed a lower value for the stress gradient for the grazed scenario (C1), and the calibration selected a more water-saving (lower Emax and/or gmin) strategy. These differences were counteracted to some extent by an increased investment in aboveground biomass (higher lmro) and a more resource-exploitative strategy (higher SLA). The C PFT showed similar differences except for the reduction of the minimum canopy conductance (gmin). However, this is likely an artefact of the parameterisation. As stated in Sect. , both SLA and lmro not only underpin the compensation of defoliation but also play a role for resource uptake and distribution. In the ungrazed scenario (C0), no defoliation has to be compensated and both parameters only play a role for resource uptake and distribution, which likely affected the selection of gmin. In contrast in the grazed scenario (C1), gmin and Emax become more important for resource uptake and distribution. For the disturbance gradient, all PFTs had higher values from different causes: the C PFT established less offspring (lower kest), the S PFT increased its stature (higher kbeer) and seedling size (higher LAIsapl), and the R PFT only increased its seedling size (higher LAIsapl).

Consistent with the findings for the other sites, for the cold steppe all PFTs showed different strategies for the different management intensities. While the value for the stress gradient was the same for the R PFT and only differed for the C and S PFTs, the calibration selected different trait values for all PFTs. For the C PFT, the calibration selected a less water-saving strategy (higher Emax and gmin) and for the S PFT a more water-saving strategy (lower Emax and gmin). For the disturbance gradient, the C and S PFT showed a higher value, and the R PFT showed a lower value in the intensively grazed scenario (S6). While for the C PFT this was the result of an increase in the efficiency of its root network (higher kroot), for the S PFT this was a result of an increase in stature (higher kbeer; Sect. ). In contrast, the R PFT had a smaller stature (lower kbeer) and seedling size (lower LAIsapl).

Across sites, we found a large variation within both dimensions which ranged from 0.30 to 0.64 for the stress and from 0.18 to 0.68 for the disturbance gradient (Fig. ). As a consequence of our assumptions for the parameterisation, the sorting of the parameter values for the three PFTs had to match the hierarchy defined in Table  (Sect. ) for each site. Between sites however, we did not make any assumptions that would predetermine an order, meaning that each site could occupy a different area of the two dimensions. For example, an R PFT had to have a higher value for the stress gradient compared to the S PFT for the same site, but could have a lower value compared to the S PFT of another site, as is the case when comparing the temperate grassland to the hot steppe. For the disturbance gradient, the same case can be made.

However, if averaged over all sites and management scenarios, the C PFT still was the most resource exploitative with a value of 0.55 for the stress gradient, while the R and S PFT were more resource conservative with values of 0.48 and 0.25. Similarly, the R PFT produced most offspring and had the smallest stature with a value of 0.29 compared to 0.57 and 0.58 for the C and S PFTs. While this general pattern emerged clearly for the two dimensions, there were substantial differences between the sites when comparing the contributing parameters. Most similar was the lmro determining investments into aboveground versus belowground biomass, which contributed to high values of the C and R PFTs for the stress gradient for several scenarios. For the steppe sites, there was some alignment within the S PFTs, which all had a larger stature (higher kbeer). The remaining parameters were not discernibly aligned across sites.

The temperate grassland already is a productive site where water and nitrogen are not limiting productivity, and we did not simulate any additional scenarios but focused on comparing the two fertilisation levels (N0 and N1). For both scenarios, total annual forage offtake was similar and between 5.3 and 7.4 Mg DM ha-1 yr-1 for the unfertilised and between 4.7 and 8.9 Mg DM ha-1 yr-1 for the fertilised scenario (Fig. a). The first cut was the most productive, yielding between 1.8 and 2.8 Mg DM ha-1 yr-1 for the unfertilised and between 2.5 and 3.9 Mg DM ha-1 yr-1 for the fertilised scenario. The subsequent cuts contributed substantially to the overall forage offtake except in 2018, which was a drought year. Here, the forage offtake from all cuts was reduced. In all cuts, the dominant C PFT contributed the majority of the forage offtake. In both scenarios, the S and R PFTs barely contributed (2 % and 8 % share of forage offtake on average) in all cuts (Fig. S8). Overall, dry matter yield (DMY) was more stable between years (except 2018) in the fertilised scenario because of higher yields during the regrowth stages (cuts 2 to 4). These compensated the slightly lower DMY of the first cut compared to the unfertilised scenario.

The SOC showed no significant trend for the unfertilised scenario, where the annual average decreased by 0.02 Mg C ha-1 yr-1 on average (τ=0.09, p value 0.1). In contrast, SOC in the fertilised scenario increased by 0.96 Mg C ha-1 yr-1 on average (τ = 0.56, p value < 0.001), respectively (Fig. d). Intra-annual SOC dynamics, which are driven by the litter production and C input from manure, were stronger in the fertilised scenario.

For the hot steppe, we simulated an irrigated (I) scenario in addition to the rainfed (R) scenario which was used for the calibration of our PFTs for the ungrazed (C0) and grazed (C1) management. Annual forage offtake was 0.26 Mg DM ha-1 yr-1 in the rainfed scenario (Fig. b), and animal feed demand was always met (Fig. S9a). Similarly, the feed demand was always met in the irrigated scenario (Fig. S9b). However, between the two scenarios the composition of forage offtake strongly differed. In the rainfed scenario, the S PFT contributed the majority in most years, whereas in the irrigated scenario the community composition changed, and all PFTs contributed to forage offtake similarly. A shift also occurred in the ungrazed scenario, which was still dominated by the S PFT but showed a higher share of the C and R PFTs after several years as well. This change was related to the changing community composition (Sect. ) and increased leaf biomass in the irrigated scenario. In the ungrazed scenario, 55 % of the leaf biomass increase from irrigation resulted from elevated growth of the S PFT, 21 % from the C PFT and 23 % from the R PFT. This was different in the grazed scenario, with -18 % (S PFT), 82 % (C PFT) and 37 % (R PFT), respectively.

The SOC of the rainfed scenarios did not show strong trends (Fig. e). However, the negative trend in the ungrazed scenario (C0) was still significant (τ=-0.27, p value < 0.001). In the irrigated scenario, SOC increased strongly with little differences between the grazing scenarios – on average by 4.9 Mg C ha-1 yr-1 in the ungrazed and 4.2 Mg C ha-1 yr-1 in the grazed scenario. However, SOC did not increase linearly but showed a much stronger increase which was unrealistically high in the first 1 to 2 years after the start of irrigation (10.4 Mg C ha-1 yr-1 in the ungrazed and 10.6 Mg C ha-1 yr-1 in the grazed scenario) than in the remaining time series (2.1 Mg C ha-1 yr-1 in the ungrazed and 1.0 Mg C ha-1 yr-1 in the grazed scenario).

For the cold steppe, we simulated an irrigated (I) and a fertilised (F) scenario in addition to the rainfed (R) scenario used for the parameterisation for both the low (S1) and high (S6) grazing intensities. Total forage offtake was 0.17 Mg DM ha-1 yr-1 for all scenarios with low grazing intensity because the feed demand of the animals was always met (Fig. c). In all scenarios, the forage offtake was almost entirely attributed to the dominant S PFT, (Fig. S10a–c). For the high grazing intensity, total forage offtake was 1.03 Mg DM ha-1 yr-1 if the feed demand was met. This was always the case in the irrigated scenario but not in the rainfed and fertilised scenarios. In the latter two, the model simulated very similar forage offtake, indicating that nitrogen addition was not sufficient to increase productivity because water was the main limiting factor. In all three scenarios, the S PFT was dominant (Fig. S10d–f). However, in the rainfed and fertilised scenarios the share of the S PFT decreased in months when the feed demand could not be met and mainly the share of the C PFT increased. In the irrigated scenario, only the S PFT contributed to the forage offtake (for an explanation, see Sect. ).

SOC was similar for the rainfed and fertilised but differed for the irrigated low- and high-grazing-intensity scenarios (Fig. f). Both the rainfed and fertilised scenarios showed a significant negative trend for SOC, which was similar between the high grazing intensity where SOC decreased by roughly 4 Mg C ha-1 yr-1 on average (τ=-0.99, p value < 0.001) and the low grazing intensity with SOC losses of 3 Mg C ha-1 yr-1 on average (τ=-0.87, p value < 0.001). For the irrigated scenarios, SOC increased by 2.5 Mg C ha-1 yr-1 on average (τ=0.79, p value < 0.001) for low grazing intensity and 0.4 Mg C ha-1 yr-1 on average (τ=0.47, p value < 0.001) for high grazing intensity.

Each site showed substantial intra-annual dynamics of total aboveground biomass (Fig. S11a, b, S12a, c, S13a, g) and the monthly average of the aboveground biomass share of the C, S and R PFTs (Fig. ). However, the intra-annual dynamics were different between sites. In the temperate grassland, the C PFT was dominant throughout the year; however, after the end of a growing season, the marginal PFTs had an increasing share until after the first cut (Fig. S11c, d). While in the unfertilised (N0) scenario the share of the S PFT increased, the share of the S and R PFTs increased in the fertilised (N1) scenario (Fig. a).

In the hot steppe, the community was dominated by the S PFT in both management scenarios (Fig. b and d). In the ungrazed (C0) scenario, the C PFT made up almost the entire remainder of the aboveground biomass (Fig. S12a). However, the C PFT was replaced by the R PFT in the grazed (C1) scenario (Fig. S12c).

For the cold steppe, PFT shares of aboveground biomass did not show strong intra-annual variation for the extensive (S1) grazing scenario (Fig. c). However, for the intensive (S6) grazing scenario, the C and R PFTs strongly contributed to the overall leaf biomass, and the C PFT was even dominant during and after the grazing period (Fig. S13h).

Removing resource limitations led to a shift in the community composition for the hot and cold steppe.

The hot steppe transitioned from an S-dominated community to a community with more balanced CSR shares that was still dominated by the S PFT (Fig. b and d). This transition occurred within the first 1 to 2 years after the beginning of irrigation for both scenarios (Fig. S12e–g), which was reflected through the shift in the community average in Fig. b and d. Whether or not this is the new equilibrium state or the community is still transitioning is crucial (Sect. ).

While removing the nitrogen limitation did not alter the community composition of the cold steppe under extensive (S1) and intensive (S6) grazing, irrigation had an effect (Fig. c and e). The S PFT out-competed the other PFTs entirely in the both grazing scenarios throughout the time series (Figs. c and e and S13e, f, k, l).

While the fertilised scenario for the temperate grassland was already well simulated in LPJmL 5, the unfertilised scenario underestimated forage offtake (Sect. ). In LPJmL-CSR, growth of the vegetation was faster than in LPJmL 5 which led to higher yields for all cuts. We identified two reasons for the faster growth. First, the new implementation for biological nitrogen fixation (Appendix ) reduced nitrogen stress and promoted higher photosynthesis rates. Second, while the parameters used for LPJmL-CSR were tuned for performance under the site-specific environmental conditions and management, the parameters used in LPJmL 5 were defined for large-scale simulations with different management scenarios.

The temperate grassland is neither water nor nutrient limited, and since we only assessed scenarios with reduced resource limitations, we only compared the fertilised and unfertilised scenarios. Despite the additional nitrogen input in the fertilised scenario, the unfertilised scenario achieved a similar forage offtake. Missing nutrients were acquired through biological nitrogen fixation, which was much higher in the unfertilised scenario, which is in line with the higher share of legumes observed in the field experiments . Despite the higher share of legumes in the unfertilised experiments, the share of C, S and R strategists was similar, and both fertilisation levels were dominated by C species, which was well-represented by the model.

The simulated SOC was strongly dependent on the land use history for which available data were limited. For simplicity, we did not simulate crop rotations for the land use history but selected a livestock density of 1.0 LSUs ha-1 (where LSU represents livestock unit) for the land use spin-up simulation (see Sect.  and the Supplement) to prescribe a fixed grazing pressure, which led to an underestimation of observations in the unfertilised scenario in LPJmL 5. This indicated that carbon inputs into the soil were too low in LPJmL 5. LPJmL-CSR showed smaller deviations from observations and an adequate representation of the trends (Sect. ). The increased soil carbon input had three reasons. First, the trade-off between SLA and leaf longevity led to higher turnover rates and in turn higher litterfall compared to LPJmL 5. Second, accounting for mortality explicitly constituted an additional input into the litter layer. Third, our simulation included manure application which provided an additional carbon input into the system.

The community composition showed some intra-annual variability, and higher shares of the marginal PFTs at the end and the beginning of a growing season in the unfertilised and fertilised scenarios (Sect. ). The S PFT gained higher shares in the unfertilised scenario, showing an advantage of the S over the R PFT despite the fact that strong nitrogen stress was avoided through biological nitrogen fixation. In contrast, if nitrogen stress was removed entirely, the S PFT lost its advantage and the R PFT could increase its share. After the first cut, these shares of the S and R PFTs became smaller, because a cut is a disturbance that directly removes part of the aboveground biomass. One strategy to cope with this is grazing (or in this case mowing) tolerance , which requires fast regrowth of the leaves to compensate for the removed biomass as is typical for a C strategist (, and Sect. ).

For the hot steppe, LPJmL 5 performed better for SOC, while LPJmL-CSR performed better for forage offtake. We identified several reasons for these inconsistent results. First, LPJmL does not distinguish between leaves of different age classes and therefore not between alive, senescent or moribund tissue . All tissue is either alive and associated with the plant or moribund and part of the litter layer. However, observed forage offtake also included senescent biomass . This predisposed the model to underestimate forage offtake when accounting for realistic turnover rates, which was observed in the low biomass values simulated in LPJmL-CSR. Second, litter decomposition is a function of soil moisture, temperature and litter composition . However, the PFTs do not differ in their persistence of the litter, which is the case for different plant species and across ecological strategies . Considering this may help to improve the simulation of SOC dynamics in the future. Third, the vegetation was described as an open thornbush savanna which includes a woody component. However, in LPJmL managed grassland vegetation does not include bushes or trees and therefore only partially represents the observed community.

The S PFT was dominant in the grazed and ungrazed scenarios, while the remainder of the aboveground biomass was contributed by different PFTs depending on the scenario (Sect. ). The dominance of the S PFT independent of grazing is plausible considering the pronounced dry vs. wet season dynamics at the site that impose water stress and potentially also nitrogen stress. The R PFT was more tolerant towards grazing disturbances and gained dominance in the grazed scenario, replacing the C PFT which had a lower ability to deal with disturbance. Removing the water limitation led to an increase in forage offtake and SOC, which can be expected when removing the main resource limitation. However, the majority of the SOC increase occurred in the first 2 years after the start of irrigation, which is not realistic. This can be explained by the missing representation of senescent tissue in combination with the adaptation of the community composition: removing the water limitation led to a strong increase in leaf biomass, which was substantially higher than the feed demand of the simulated grazing intensity and increased the input to the litter layer. Furthermore, the share of the R and C PFTs which have a lower leaf longevity than the S PFT increased, leading to faster inputs into the litter layer. After 1 to 2 years, the community composition reached a new equilibrium, and inputs into the litter layer decreased. Introducing senescent tissue would increase the competition for light due to self-shading effects and likely slow down this transition.

In addition, irrigation led to a shift in the community composition (Sect. ) and an increase in leaf biomass to which the C and R PFT together contributed more than the S PFT (Sect. ). We cannot determine whether or not increases under irrigation would be lower for an S-PFT monoculture, which does not contain other ecological strategies, but we strongly suspect so.

In both the ungrazed and the grazed scenario, the community transitioned from strongly S-dominated to a community with higher shares of the C and R PFTs that were still S-dominated. According to the CSR theory, this type of community emerges in somewhat stressed and disturbed habitats . While this case can easily be made for the grazed scenario, where the disturbance is caused by the animals, the ungrazed scenario does not include such a clear disturbance. The success of both the C and the R PFTs is likely determined by the similarity of their SLA, kbeer and lmro, which become more important compared to Emax and gmin if there is no water limitation. Potentially larger differences in these parameter would lead to the success of one of the two instead.

Less than 2 years is a very fast transition, and while the shares of the leaf biomass seem to have reached a new equilibrium after 1 or 2 years of irrigation, it is likely that the soil carbon and nitrogen pools are not in equilibrium yet. This is especially interesting when considering that the overall increase in leaf biomass may promote litterfall and the formation of inorganic nitrogen. This in turn may lead to reduced nitrogen limitation and additional changes in the community composition. Furthermore, biological nitrogen fixation is dependent on soil moisture and may therefore also contribute to decreasing nitrogen stress under irrigation. However, irrigation also leads to increased leaching and could therefore also decrease inorganic nitrogen availability. Future analysis considering longer timescales may help to identify intermediate and final transition states.

Regardless of the equilibrium state of the transition, its velocity is likely overestimated by LPJmL for two reasons. First, the C and R PFTs can establish quickly despite their limited presence before the onset of irrigation because LPJmL does not simulate a seed bank which would in reality be small at least for the C PFT limiting its establishment. Second, in reality growth of established individuals is limited and a transition as simulated is strongly controlled by reproduction and dispersal, which slow down population biomass increase. In LPJmL, already established individuals continue to grow and the population biomass increases even without additional establishment.

LPJmL 5 underestimated the observed forage offtake of the cold steppe, because the feed demand, which was originally designed to represent large cattle , was scaled down linearly with animal body weight. This led to an unrealistically low feed demand because the feed demand body weight relationship is not linear but follows a power law . Our new calculation of feed demand (Appendix ) led to a higher feed demand and forage offtake simulations were improved for low and high grazing intensities.

Under observed conditions, the high grazing intensity severely reduced aboveground biomass, and feed demand was not met in all years except the year directly after the increase in stocking density, indicating overgrazing. The reduced biomass availability was also observed by in their field experiment. Additionally, LPJmL simulates a different community composition compared to the low grazing intensity. The relative share of the C and to some extent also the R PFT is higher for the high grazing intensity (Figs. S10b and S13h) because such strategies are better suited to tolerate grazing.

During and after the grazing period, the C and R PFTs had a higher share of the community aboveground biomass. Both these PFTs can regrow faster and invest more into aboveground biomass, which gave them an advantage over the S PFT under grazing. In addition to the observed environmental conditions, we simulated two scenarios where we removed the water and nitrogen limitations separately. Removing the nitrogen limitation barely affected biomass availability, and forage offtake was similar compared to the rainfed scenario (Sect. ). The additional soil nitrogen could not be utilised by the plants, because water was the main limiting factor . In contrast, removing the water limitation led to an increase in leaf biomass, and forage offtake met the demand in all years even for the high grazing intensity. This is in line with irrigation and fertilisation experiments conducted in the cold steppe and other sites with similar conditions e.g.. Contrary to the results of , who reported a lower share of annuals and bi-annuals – that are more likely C than S strategists – in the rainfed treatments, the S PFT was dominant in the irrigated scenarios. One reason for this could be that LPJmL does not simulate seed banks, which play a major role for the establishment and success of the annuals and bi-annuals . Instead, LPJmL simulates establishment of additional seedlings dependent on available space, assuming that resources for reproduction are available at any time and not dependent on past investments into seed production.

Despite the fact that we did not have separate data on SOC under the two grazing intensities, our results showed a lower SOC storage for the high grazing intensity typical for overgrazed steppes e.g. compared to the low grazing intensity which constituted the typical livestock density for the region . investigated the effect of high grazing intensities on the SOC, observing significant SOC losses within 3 years of increased grazing, which is in line with our simulation results. Fertilisation had no effect on SOC, because leaf biomass and in turn carbon inputs into the soil did not increase. In contrast, irrigation led to an increase of SOC, which was stronger for the low grazing intensity. This is likely because more biomass was produced, and the surplus of the feed demand was not removed but contributes to the litter layer. However, these gains would not justify the effort that would be necessary to irrigate large areas.

Removing the water limitation led to a transition from S-dominated to an S-monoculture community under both grazing intensities (Sect. ). Since the site was still severely nutrient-limited and exposed to low temperatures, it seems that an S strategy remained advantageous. Furthermore, the S PFT showed trait values associated with large investments in roots and more persistent root tissue (Sect. ), which provides a likely explanation for its increased dominance: it had an advantage in the competition for the additional water. Similar to the hot steppe, it is possible that our time frame is too short for the soil pools to have reached a new equilibrium. As described in Sect. , irrigation alone already affects processes that could increase nitrogen supply by biological nitrogen fixation and litterfall, but it could also decrease it by leaching. Both biological nitrogen fixation and mineralisation are dependent on soil moisture as well as on temperature, which is low in the cold steppe limiting the increase of inorganic nitrogen. Therefore, it is possible that only an intermediate state emerges during our simulation period. Especially when also considering the increased leaching, we expect that the cold steppe is still nitrogen limited under irrigation; therefore, combining irrigation with fertilisation could further reduce nitrogen limitation, leading to increased productivity and changes in the community composition. However, the leaf biomass increase may also be limited by higher maintenance respiration which is connected to leaf nitrogen content. Additional analysis is needed to enhance the understanding of these complex interactions.

We used a Bayesian calibration method to find suitable parameter values of eight parameters assigned to two trade-off dimensions for the new PFTs. Due to lacking data on starting values and ranges for the three new PFTs, we used the same ranges and starting values for each PFT but prescribed an order of the parameters. Within a site and management scenario, the prescribed hierarchy for specific parameters also predefined the ranking of the PFTs along the stress and disturbance gradients. Across sites and management, we did not constrain the PFTs to positions within the two dimensions. Theoretically, all PFTs of the temperate grassland could have been associated with a more conservative strategy for the stress gradient compared to the PFTs of the hot steppe. However, while there were some differences between the sites and management; on average, the C and R PFTs occupied a more resource-exploitative position for the stress gradient and the S PFTs a more conservative one (Sect. ). Similarly, for the disturbance gradient the C and S PFTs occupied a position associated with less but larger offspring and a larger stature compared to the R PFT. It is an emergent property of the model that not only the relative position of the PFTs of a site and management scenario determined community composition but also the overall positions along the stress and disturbance gradients (which we derived from the global spectrum of plant form and function ) were important. Our experiences from these three sites showed similar strategies that are independent of environmental conditions, indicating that LPJmL-CSR is capable of reproducing the empirically derived trade-offs associated with the global spectrum of plant form and function . However, LPJmL-CSR will benefit from additional testing on larger scales in the future.

While missing processes such as the representation of seed banks as at the hot steppe (Sect. ) and poor data as at the cold steppe (Sect. ) may have led to biased model dynamics to some extent, we clearly demonstrated the importance of representing different ecological strategies.

The calibration selected different strategies along the stress and disturbance gradients for the different management intensities (Sect. ), which were related to changes in resource limitations or disturbance level: in LPJmL-CSR, a change in resource availability only changes the conditions for the establishment of a community but does not directly affect the established vegetation changes in environmental filters;. In reality, however, a change in resource availability may also increase the mortality for specific strategy types, affecting the already established community as well. In temperate grasslands, manure application increases N supply and reduces the number of available niches that can be occupied by different ecological strategies. In the unfertilised experiment, species could satisfy their N demand through two different strategies: competition for the limited resource in the soil or biological N fixation (BNF). In the fertilised experiment, only the first strategy was advantageous as BNF creates additional costs. In the field experiment, this was evidenced through the substantially different amounts of legumes between the two experiments . In the model, N-fixing and non-N-fixing species are both collated within each PFT. Therefore, in the unfertilised scenario, a PFT had to apply a strategy combining N uptake and fixation, whereas it could focus on N uptake in the fertilised scenario. Since we calibrated the unfertilised and fertilised scenarios separately using the same data for C, S and R PFT covers, the difference in strategy between the two scenarios is expressed through the different positions of the PFTs along the stress and disturbance gradients: higher investments into belowground biomass (lmro) provide an advantage in the competition for plant-available nitrogen . In the model, this led to a reduced need of fixing additional nitrogen and in turn a reduction of the investment costs associated with biological nitrogen fixation (Sect. ).

In contrast to resource availability, a disturbance directly affects the vegetation. In the case of grazing, it also influences resource availability indirectly through removal of nutrients from and spatial redistribution within the system . In LPJmL, the grazing of the animals at the steppe sites constituted a direct reduction of leaf biomass proportional to the cover of each PFT . Under intensive grazing, strategies of grazing tolerance or avoidance are essential . While grazing tolerance is mainly associated with fast regrowth (, stress gradient), grazing avoidance strategies can operate in time and space. Grazing avoidance in time is possible through the completion of the lifecycle between grazing intervals (, disturbance gradient). Grazing avoidance in space is contingent on reducing plant size . However, since plant size is not explicitly represented in LPJmL, we do not discuss this strategy further (Sect. ). In the hot steppe, we simulated a daily grazing system, which makes grazing avoidance through the lifecycle impossible, and the PFTs had to follow a grazing-tolerance strategy. This was expressed through changes in the stress gradient: all PFTs increased their investment into aboveground biomass and faster tissue growth (Sect. ). Because LPJmL does not account for differences in the palatability of different strategy types the parameterisation could not select for such likely successful strategies, leading to a potentially biased community composition.

At the cold steppe site, grazing only happened during the growing season, and both grazing tolerance and avoidance could be useful strategies. However, grazing avoidance in time, which is the only type simulated by LPJmL, will not be successful as it would mean shifting biomass production to the non-growing season where the environmental conditions do not allow growth. Still, between the extensive and intensive grazing scenarios, the differences between the PFTs in both dimensions do support different strategy adjustments (Sect. ). The C PFT increased its investment into aboveground biomass to tolerate grazing, while the S and R PFTs did not show any adjustment. However, since the high grazing pressure caused degradation of the aboveground biomass, differences between the two management scenarios not only reflect different strategies to deal with the disturbance but also reflect different strategies for survival outside the grazed period. As such, all PFTs constructed long-living tissue to survive unproductive conditions outside the growing season in the intensive grazing scenario. This was not necessary in the extensive grazing scenario, because the PFTs retained substantial aboveground biomass at the end of the growing season and did not need to be as resource conservative.