The QUINCY terrestrial biosphere model simulates the coupled carbon, nitrogen (N) and phosphorus (P), as well as water and energy cycles. It represents physical processes relevant to high latitudes, such as snow dynamics, soil freeze-thaw, and adaptive rooting depths to layers in non-frozen layers . The standard version of the model includes 12 different PFTs with flexible parameterizations for leaf characteristics, phenology, and other growth and turnover dynamics. The concept of PFTs groups all plant species into a limited number of classes, thus allowing for a much more traceable interpretation of the plant interactions . The approximately 40 parameterizations representing the different plant characteristics and functionalities in QUINCY are based on plant trait databases . Here, we have added additional PFTs to simulate Arctic shrubs.

In QUINCY, vegetation consists of structural pools (leaves, sapwood, heartwood, coarse roots, fine roots, and fruit) and a fast-overturning, respiring non-structural pool (labile), as well as a seasonal, non-respiring, and non-structural storage pool (reserve). The canopy comprises several layers (a multilayer canopy scheme), including separate representations of photosynthesis and canopy conductance for sunlit and shaded leaves separately within each layer. Vegetation processes as well as nutrient uptake and availability respond to local environmental conditions on a process-specific timescale, dynamically enabling a certain memory effect . Plants partition the required C between growth, respiration and storage, acclimatising to meteorological conditions and nutrient availability.

The soil is explicitly discretized vertically into 15 layers of increasing thickness down to a depth of 9.5 m, and this applies to soil temperature, texture, moisture, and the biogeochemical soil pools. Soil biogeochemical pools are represented by five organic pools: structural; polymeric; and woody litter; and fast and slow-degrading soil organic matter; as well as solute concentrations of ammonium (NH4), nitrate (NO3) and phosphorus (PO4). Each of these organic pools contains C, N, and P . The model specially accounts for the nitrogen cycle in soil by simulating nitrogen inputs through symbiotic (microbial) fixation in higher plants, as well as free-living (non-symbiotic) N fixation in all parts of the ecosystem. It also accounts for N deposition, and the transformation processes of mineralization, nitrification, and denitrification, as well as the transport between soil layers. N is always accounted for as NO3- and NH4+.

Site-level QUINCY model simulations were conducted for sixteen Arctic sites (Fig.  and Table S1 in the Supplement) from 1901 to 2010. Data on shrubs and GPP estimates were provided by the ABCflux dataset version 2 for these sites. In the reference simulation, each site is represented by one of the two shrub PFTs according to the vegetation reported at each site (see Table S1). The simulations include nitrogen dynamics, as mineral nitrogen is often known to limit plant growth in the Arctic . Different model configurations were chosen to quantify the differences between ecosystems covered by graminoids and NE trees at the same sites. A second set of simulations disentangles the effects caused by climate drivers, CO2 fertilization and nutrient limitation (Table ).

The standard initialization of QUINCY does not account for the elevated levels of soil organic C and N characteristic of permafrost soils. These contents cannot be built up in a standard QUINCY spin-up either, as this would require a simulation of the type used for paleo-timescales that reproduces at least the last interglacial and glacial periods. This is neither practical for our purposes, nor feasible given the lack of the required model forcing data. In this study, we therefore adapted the initialization procedure by adjusting the C and N contents, as well as the distribution of the initialization, when a site is identified as permafrost based on the definition and methods in . The permafrost distribution curve is initialized using a Weibull function with a shape parameter of 0.6 instead of 2, and the minimum relative deep permafrost concentration is set at 0.3 of the surface soil concentration according to the estimated distribution. Soil organic matter is initialized with a single permafrost content of 121 000 gCm-2yr-1 for the first 3 m, which is the vertically integrated average C. Permafrost nitrogen is initialized using a C:N ratio of 14. It should be noted that the applied approach does not account for spatial heterogeneity in permafrost soil content and distribution during initialization.

The meteorological forcing that drives the model is derived from the University of East Anglia Climatic Research Unit Japanese Reanalysis (CRU-JRA; ), which is a reanalysis of incoming solar radiation, surface temperature, humidity, precipitation, surface pressure, and surface winds. This dataset is available with a spatial resolution of 0.5°×0.5° and at a temporal resolution of 6 h. Data for our selected sites were extracted from the CRU-JRA dataset according to their geographic coordinates . Furthermore, the soil is represented by 15 layers, with increasing thickness down to a depth of 9.5 m. The model starts the transient simulations after a spin-up period of 1000 years using cyclic forcing data from 1901–1930. The atmospheric CO2 concentrations are prescribed according to .

We implemented two new shrub PFTs: broadleaved deciduous (BD) and needleleaf evergreen (NE). The two shrubs are based on the respective default parametrization of the woody PFT, but due to the physiological and phenological differences important aspects concerning the initial leaf N, plant allometry, root distribution and wood turnover times are specified based on measurement studies and trait databases (e.g. TRY, ). The new parameters are listed in Table .

The ABCfluxnet v2 data, cited in the references , , was recently compiled to provide a dataset of in situ Arctic–boreal terrestrial ecosystem CO2 flux measurements. The synthesized GPP is given as cumulative flux densities aggregated at the monthly timescale. These methods were employed during the growing season. We selected sites where shrub growth has been observed at multiple points in time. The 16 study sites are located across the northern USA, Canada, Finland, Norway, Sweden and Siberia, as shown in Fig. . These sites are in wet, moist or dry tundra, fens, bogs or boreal forests and are vegetated by either a mixture of shrubs, trees and grasses or only shrubs (see Table S1). The measurements were averaged and used across different years to minimize the impact of climate variability. The data set includes eddy-covariance (EC) measurements, which generally represent a 100 m radius footprint of the measurement tower. For some sites, only chamber data exists (Table S1) which can be considered the most highly resolved, plot-scale GPP estimate of a specific vegetation/soil unit.To further evaluate the simulated GPP, we also use data extracted from the global MODIS/Terra Level-4 products of MOD17A2H (Collection 6) and MOD17A2HGF (Collection 6.1) at the site coordinates. Both products provide a cumulative eight-day composite of GPP values at a spatial resolution of 500 m and have been interpolated. GPP is estimated by applying a radiation-use efficiency framework that integrates absorbed photosynthetically active radiation (APAR) with the biome-specific radiation conversion efficiency (ε), which is constrained by meteorological stress. While the MOD17A2H product contains gaps due to the quality of the input data, the MOD17A2HGF product reduces data gaps by employing additional temporal gap-filling based on quality control flags. Both products are subject to uncertainties relating to the quality of the input data, land cover classification, and model assumptions. With a GPP time series that is more continuous in space and time, the MOD17A2HGF dataset introduces an additional source of uncertainty in the form of interpolation, which dampens the signals caused by inter-annual variability in climate or ecosystem disturbances. Extensive validation against eddy covariance measurements shows that MOD17 series products can moderately capture the interannual and seasonal variability in GPP, with R2 over 0.6 across different biomes and climate regimes . We extracted GPP time series from the datasets for the different sites using the nearest neighbour approach.

The modelled shrubs capture a reasonable monthly mean seasonal cycle of GPP at half of the validated sites (six out of twelve, with a 50 % measured seasonal cycle), when compared to the field measurements in Fig. . Modelling with the new NE shrub PFT improves the mean R2 value to 0.46 across nine sites, compared to the default (tree) PFT (see the dashed line in Fig.  and Table S2 in the Supplement). In contrast, simulations with the BD shrub (7 sites) mostly underestimate the measured seasonal cycle of GPP; the default (tree) PFT therefore agrees better with the measurements (mean R2=0.6). For further evaluation, we compared our model's GPP outputs to GPP MODIS satellite products. MODIS data is widely used to evaluate the carbon cycle in global terrestrial biosphere models and in the Coupled Model Intercomparison Project (CMIP) models (50 % use the JSBACH surface scheme) , showing a high agreement in the high latitudes. In this study, the MODIS GPP products show a higher correlation with the model, with mean R2 values of 0.71 for NE shrubs and 0.50 for BD shrubs (Table S2) than compared to the site measurements. NE shrubs generally improved the agreement with the MODIS data compared to the NE default (tree) PFT (R2=0.59). The MODIS products generally align with field measurements, though they diverge significantly at two sites. At the ISK site (Norway), the GPP of MODIS and the (CN) model agree quite well, whereas the GPP measured in the field is much larger. This may be due to the specifics of the local climate and landscape at the site coordinates used. Specially, half of the field measurements at the ISK site are taken in wet tundra areas whereas however the model soil hydrology tends to be rather dry. At the LET (Finland) site, which has a significant proportion of forest (e.g. ) both observational datasets show a higher GPP which the model does not reflect because it represents shrubs only (in this simulation). Another possible source of the model bias is the underestimation of soil moisture at sites with a higher wetness than 25 %, such as the Stordalen (STR, Sweden) site . Overall, two out of sixteen sites (Lettuoso (LET) and Stordalen (STR)) are considered not to be representative of the simulated shrub PFTs due to their heterogeneous ecosystems.

All in all, the model's performance depends on the chosen configuration, the evaluated shrub PFT, local heterogeneity of vegetation coverage. While this comparison is an important benchmark for the model and its new parametrization, it is limited to monthly GPP values. The structural and phenological plant characteristics can be only discussed quantitatively in the later analysis of this study due to limited data availability and time constraints.

The simulated (multi-site) average annual GPP flux at shrub sites is 371 gCm-2yr-1, ranging from 132 to 735 gCm-2yr-1 for the time period 1990–2010 (Fig. a). These estimates (and their partitioning) are consistent with the data synthesized by who report an average (area-upscaled) Arctic GPP of 300 gCm-2yr-1 for tundra and 416 gCm-2yr-1 for the Arctic permafrost region, although this accounts for all tundra vegetation species (graminoids, shrubs, mosses, and lichen). The annual GPP for shrubs has a much wider interquantile range than for tress and graminoids, i.e. it is more variable (Fig. b). The distribution is more symmetrical, with shorter tails and a flatter peak and a lower skewness and kurtosis. This shows the better adaptation of shrubs to the cold Arctic climate, as it has been observed elsewhere e.g.. Thus, the median (not considering the high values) of the new shrub and tree PFTs are not significantly different, while the average annual GPP increases by 16 % due to the implementation of shrubs. Additionally, report a substantially higher GPP at shrubs than at trees when simulating a Canadian site with the CLASSIC model, which aligns well with the site measurements. Shrub dominance instead of graminoid cover would increase the annual mean and median GPP by 40 %, respectively.

The GPP change of the three PFTs is determined by different N pathways, and the strength of the correlation varies not only across the PFTs, but also between sites. For example, the GPP at NE trees and shrubs increase much more strongly in response to NH4+ uptake than at graminoids, whereas comparing graminoids and trees to BD shrubs reveals a similar slope (Fig. S1 in the Supplement). Thus, among the shrub species, the model simulates faster growth in response to NH4+ uptake for deciduous shrubs, as also reported by . For all shrubs, the change in NH4+ uptake explained 50 %–70 % of the GPP change compared to trees (see R2 in Fig. S2 in the Supplement). For NE shrubs, N fixation and NH4+ net mineralization also showed a significant correlation with the GPP changes (Fig. S2). Most of the increase in GPP at evergreen shrubs could be additionally explained by higher biological nitrogen fixation (BNF) compared to the default PFT. Here, the absolute numbers approximately half the magnitude of the plant NH4+ uptake (Fig. S3 in the Supplement). Indeed, and report substantially higher asymbiotic BNF in shrubs than in trees, driven by bacterial activity. The GPP changes in deciduous shrubs are usually lower than in evergreen shrubs, which may explain the low correlation. The GPP differences from graminoids are generally less correlated with changes in the N uptake pathways.

The site-averaged foliage (projection) cover of shrubs (i.e. the percentage coverage of the ground area covered by leaves when viewed from above), is very similar to that of trees (4 % higher for shrubs), indicating slightly higher self-thinning in shrubs (Fig. S8 in the Supplement).

Shrub cover reduced the carbon-use efficiency (NPP/GPP), at all sites, from 0.64 (graminoids) and 0.68 (trees) to 0.63 at NE and 0.57 at BD shrubs, indicating lower efficiency in converting GPP to biomass overall (Fig. S5a in the Supplement). This is generally consistent with satellite estimates of the NPP/GPP ratio for the three different vegetation types reported by . Unlike needleleaf evergreen shrubs, most deciduous shrubs had lower net primary production (NPP) than trees (Fig. S5a) following the growth respiration (Fig. S5b in the Supplement). Higher growth respiration indicates a faster growth rate and higher seasonal leaf turnover .

The new shrub PFTs produced reasonable total aboveground C values, ranging from 0.7 to 2.5 kgCm-2 (Fig. a, site-mean=1.3 kgCm-2) between 1990–2010. Given that approximately 50 % of C in plants is found in the aboveground part these values translate into biomass estimates consistent with measurements of 0.4–1.5 kgCm-2 .

Most of the simulated C here was stored in the living wood of the shrubs followed by the storage in coarse roots and heartwood (Fig. S4 in the Supplement). ). Shrubs produce less sapwood than trees, but have a faster production rate (Fig. S6a in the Supplement). According to e.g. this is due to the narrower and fewer xylem cells in shrubs. As explained by , deciduous shrubs grew faster than evergreen species (Fig. S6a), since their small leaf mass per unit area enabled them to intercept more light for a given investment of leaf C than evergreen shrubs. Furthermore, the modelled shrubs stored more C in coarse roots than trees, increasing the coarse root:sapwood ratio by a factor eight. This is consistent with findings by and a global dataset (including 530 shrub land sites) by . Thus, the total aboveground C storage of shrub vegetation is lower overall than that of trees (at 13 out of 16 sites, Cliff's δ=-0.39). Indeed, report that trees allocate more nutrients aboveground than shrubs and graminoids.

While the distribution of annual mean vegetation carbon at shrub sites overlaps with that of trees (Fig. a), the mean biomass at sites with shrub coverage is 17 % lower than the mean tree biomass, as indicated by the lower carbon use efficiency of shrubs (Fig. S5b). The shrub distribution has a much larger interquartile range and is more skewed towards higher values than tree distribution, which points to the higher variability of shrubs. The total vegetation C of graminoids is significantly lower (effect size: Cliff's δ=1.0) than in both woody PFTs when considering distribution, median and range, as also found by . Graminoids store relatively more C in fine roots and the labile pool (almost 40 %), but woody plants have significantly larger wood and coarse root storage.

The higher fraction of coarse root C stored by shrubs determines the increase in total soil C (Fig. a), reaching 14.3 kgCm-2 compared to other PFTs (vs. 13.2 for trees and 11.1 kgCm-2 for graminoids) at 81 % of the stations. This small (6 %) increase in soil C due to shrub coverage compared to trees is consistent with observational studies reporting greater belowground than aboveground C with shrub vegetation coverage . It should be noted that the soil type is a primary factor that can explain both increases and decreases in soil organic content, as a meta-analysis of different shrub species found globally . even report a substantial C loss by shrub expansion, which is consistent with other studies. At a landscape scale, found an increase in soil organic carbon with shrub expansion in the Arctic. Our modelling study shows that, evergreen shrubs tend to conserve more C than deciduous shrubs (first eight bars in Fig. ), as reported by . In comparison with graminoids, soil C increases by 18 % as substantially more C is stored in coarse roots of shrubs. However, since the model does not yet consider the long-term C storage in permafrost soils, the impact on soil C is only discussed briefly here. Given that shrubs exhibit higher root nutrient absorption rates than trees, as supported by , who analyzed 170 root samples from Arctic shrubs, and as corroborated by our simulations (Fig. S6b in the Supplement) significant impacts on nitrogen cycling can also be expected . Overall, we demonstrate that the newly modelled shrubs produce reasonable C fluxes and storage while reproducing important physiological processes (e.g. wood formation and root allocation) that differ from those of trees.

Here, we compare the C-only (ShrubsConly) and CN (Shrubs) simulations, i.e. simulations in which the nitrogen supply is maintained at levels that do not constrain plant growth vs. an explicit simulation of nitrogen input and output processes, as well as the limitation of plant growth due to nitrogen availability. Accounting for nitrogen limitation significantly reduced the GPP at all sites in shrub cover simulations (Figs. b and S7 in the Supplement), most notably for deciduous shrubs. These reductions ranged from 3 gCmonth-1m-2 (2 %, not significant) for the evergreen shrub at the YAK site in Russia to 150 gCmonth-1m-2 (75 %) for the deciduous shrub at the IGA site in Russia (Fig. a). We demonstrate that the magnitude of GPP change (Fig. b) depends on the length of the growing season (Fig. a), which impacts the nitrogen supply to the ecosystem through net N mineralization, fixation (not shown) and uptake (Fig. c). At sites with deciduous shrubs and a three-month growing season only (at the SEI, IMN, IGA sites; whereas SEI and IGA shrubs behave very similar), these main N sources declined in the CN simulations, leading to a significant decrease in GPP (Fig. b). In contrast, shrubs at other sites, e.g. at the WLF site in Canada, showed at least an increase in N uptake at the start of the growing season under the CN configuration. When both N uptake terms and the use of N storage increase (at the sites YAK, STR and ABI, where the shrubs at STR and ABI behave similarly), the reduction in GPP is only slight (Fig. b and c). We demonstrated that nitrogen supply strongly limits most simulated shrubs, which is consistent with other studies . Other than modelled evergreen shrubs, the deciduous shrubs have a higher N demand and respond more rapidly to N additions, as has been reported for alpine shrubs by . The C-only model, which allows for unlimited biological N fixation, performs better than the CN model when compared to measurements (Table S2). This suggests that the N dynamics in the model may be overly restrictive. Supporting this hypothesis, we extracted 89 measurements of the C/N ratio at Betula nana leaves and 36 measurements of the C/N ratio at Betula glandulosa leaves were extracted from the TRY database . The C/N ratio estimated by the C-only configuration at seven stations (with deciduous species) is 24.0 (median), which is consistent with the measured median values (26.06±2.0, 22.56±2.1). However, the median value of the CN model (29.06) instead overestimates the measured C/N ratio (i.e. it underestimates the leaf N concentration), thus supporting our hypothesis of overly too strong N limitation in the CN model. We recommend to improve the model by refining key processes such as BNF, to better reflect Arctic conditions.

Higher soil N levels, originating from the improved soil initialization, stimulated primary productivity. This increases the C storage in the coarse root, sapwood and heartwood pools of the plants. Consequently, total vegetation C increased by 1 kgCm-2 (93 %) on average (Fig. S11 in the Supplement).

For most of the sites with NE shrubs, the effects of the enhanced soil N supply (Fig. S7) are clearly reflected in the annual GPP, which increased by 73–290 gCm-2yr-1 compared to the standard model simulations (Shrubs). The additional nutrients led to a significant increase in the leaf-to-root mass ratio (Fig. S13 in the Supplement), which indicates significant alleviation of nitrogen limitation. This pattern has been widely reported for other vegetation types and causes the shrubs to allocate more growth to aboveground biomass than to roots. The latter is needed to take up more nutrients. In contrast, GPP for BD shrubs is less strongly affected (Fig. S7), and the leaf-to-root mass ratio at BD shrub sites do not change significantly (Fig. S13b in the Supplement). This feature highlights the contrasting nitrogen sensitivities and acquisition strategies of NE and BD shrubs. Indeed, report a rapid increase in GPP and leaf-to-root mass ratio in NE shrubs with higher N availability whereas BD shrubs only show a slight response to increased N due to the conservative economics and higher respiration efficiency .

The change in GPP due to the PM initialization varies significantly over time. From the initialization the GPP difference rapidly increases by 10 gCm-2yr-1 to a level which is more or less constant until 1960. With increasing CO2 the GPP difference increases by 30 gCm-2yr-1 until 2010.