QUINCY is a terrestrial biosphere model that simulates the energy, water, carbon, and nutrient (nitrogen and phosphorus) cycles . For simplicity, in this study we ran the model with soil phosphorus availability set to non-limiting levels, such that plant and soil biogeochemical processes operated at average observed P contents. Alongside the nitrogen cycle, an explicit representation of the 15N cycle has been implemented, largely following the description of isotopic discrimination processes by . Different ecosystems are represented by different plant functional types (PFTs) and in the site-level version each site is represented by a single PFT.

We outline some of the basic characteristics of the model here, and a full description of the model is provided by . Plants comprise three short-lived pools (leaves, fine roots and fruits), a non-structural storage pool (reserve) and a seasonal, non-respiring pool (labile), along with three structural tissue types (sapwood, heartwood and coarse roots). All vegetation and soil pools in QUINCY contain carbon, nitrogen and 15N. Photosynthesis in the multilayered canopy is calculated following the formulation of . Photosynthetic parameters depend on the leaf N concentration . Leaf biomass and N concentration respond to soil N availability . Plant nutrient uptake depends on fine-root biomass, soil mineral nutrient concentration and plant nutrient demand.

Newly acquired carbon, nitrogen and 15N are first allocated to the labile pool, from which they can be used for tissue growth, respiration or storage in the reserve pool. The potential growth rate is influenced by the PFT-specific allometric relationships among leaves, sapwood and fine and coarse roots. Belowground carbon allocation increases under nutrient or water stress. Actual growth is further controlled by the nutrient demand of each plant pool and the nutrients available for growth, and may also be limited by air temperature and soil moisture. Flexible stoichiometry is permitted for leaves and fine roots.

Symbiotic biological nitrogen fixation (BNF) by microbial–tree partnerships is described as a dynamic trade-off between carbon and nitrogen opportunity costs, following and . Asymbiotic BNF in the litter layer depends on soil temperature, soil moisture, and the N deficit in litter decomposition. Nitrogen fixed through this mechanism is added to the fast soil organic matter pool. The parameters governing BNF were adjusted to align with the findings of .

The simulated soil has a vertically layered structure, with 15 layers of exponentially increasing soil thickness down to 9.5 m. The soil biogeochemical processes largely follow the CENTURY model approach . Each soil layer contains organic pools consisting of metabolic (soluble), structural (polymeric), and woody litter pools, fast- and slow-overturning organic matter (SOM), as well as inorganic ammonium (NH4) and nitrate (NO3) pools. The litter and SOM pool turnover rates depend on soil temperature and moisture, described by first-order kinetics. The stoichiometry of the litter pools is determined by the respective plant pool stoichiometries along with the plant resorption capacity for each nutrient. The fraction of resorption to the labile pool for leaves before shedding is 50 % and for wood 20 % during conversion of sapwood to heartwood. Fine-root mortality is assumed to be dominated by predation and therefore no nutrient resorption occurs from fine roots. The C:N stoichiometry of the fast SOM pool is affected by available inorganic N while the slow SOM pool stoichiometry is assumed to be fixed. Plants and soil microorganisms compete for the inorganic nutrients according to their nutrient demand. Nitrification and denitrification processes depend on the aerobic status, temperature, and moisture of each soil layer . As a further development of the baseline version of the model , nitrate assimilation by soil microorganisms (i.e. immobilisation of nitrate from the solute pool to the fast SOM pool) was implemented. Although this allows more rapid uptake of 15N from nitrate in the soil, this change has limited overall impact on the model results, as N immobilisation by microbial growth is primarily determined by the stoichiometric imbalance between litter and SOM, as well as by the potential litter decomposition rate in the model.

Isotopic discrimination of 15N occurs during biological nitrogen fixation, ammonification, plant and microbial N uptake, and processes associated with nitrification and denitrification. Fractionation parameters were taken from (Table ). However, the model does not represent fractionation associated with mycorrhizal-mediated N uptake. Biomass growth and N retranslocation are assumed not to fractionate N.

In this study, the model was run at the site scale, with the site represented by a summergreen broadleaf deciduous forest PFT. QUINCY uses a half-hourly time step and requires air temperature, humidity, precipitation, air pressure, short- and longwave radiation, wind speed, atmospheric CO2, 13CO2, 14CO2 mole fractions and NHx, NOy and PO4 deposition rates as forcing variables. Additional site information includes geographical coordinates and soil physical and chemical properties (texture, bulk density, rooting depth and soil depth, inorganic P content). The turnover rates of vegetation and soil pools for this PFT are listed in Table S1 in the Supplement.

We simulated the Arnot Forest using meteorological data from the CRU JRA v2.1 dataset , which were disaggregated to half-hourly resolution using a statistical weather generator . Annual atmospheric CO2 concentrations were taken from and N deposition data from . The prescribed N deposition (0.87 gNm-2yr-1; Fig. S1 in the Supplement) was similar to the observations at the site (0.7–0.9 gNm-2yr-1). The level of N deposition decreased by about one-third between 2000 and 2020, from 1.3 to 0.9 gNm-2yr-1 at a long-term monitoring site operated by the US Environmental Protection Agency, which is located close to the site . This level of N deposition is considered to represent a transition from high to moderate deposition . Soil texture and pH were taken from the observations . We did not include woody litter in the analysis, because it plays only a minor role in the δ15N cycle over the timescale considered here. We combined observations of new wood, old wood and bark and compared them with simulated sapwood and heartwood, labile, and reserve pools. Although bark can be considered an active pool in the 15N cycle , it is not explicitly represented in the model.

First, a 1000-year spinup was performed by cycling driving data between the years 1901 and 1930. After this a model simulation was performed for 1901–2018, using transient climate, CO2 concentrations and N deposition for the site. In the simulations the 15N tracer was applied as NO3 solute in the uppermost soil layer on the same dates as the tracer application was performed in the experiment. For the scenario run extending further 30 years beoynd 2018, the climatic forcing was generated using the present-day climate and the CO2 from the RCP8.5 scenario .

Additional sensitivity tests were performed to assess the extent to which the simulation results depend on the assumed fraction of N retention during leaf senescence, as well as on the turnover rate of soil organic matter pools. In these tests, we varied the coefficient determining the amount of nutrients returned to the reserve pool before leaf senescence, or increased the turnover rates of the soil pools, thereby reducing their size. As these changes affected the overall model performance without improving agreement with the observed characteristics of the carbon and nitrogen cycle or strongly affecting the temporal evolution of the tracer distribution, we used the default parameters for this PFT rather than performing rigorous parameter tuning.

First, we investigated the overall model performance in reproducing the observed characteristics of the Arnot site – vegetation and soil pools as well as C and N fluxes – because differences in pools between the observations and simulations influence the δ15N values from the tracer experiment (Table ). If a pool was overestimated in the simulations, the magnitude of the 15N signal would be biased low given observation-based flux partitioning. It is important to note that the model was not calibrated to this particular site and the default parameter values were used.

The simulated annual gaseous N fluxes were of the same order of magnitude as the observations, particularly considering the large uncertainty range in the observations. Although key indicators of productivity were comparable to the observations (wood production and leaf C, which is related to leaf area index), the model showed notable deviations from the observed partitioning of carbon and nitrogen in vegetation. The model underestimated the amount of belowground carbon, especially fine-root biomass, while generally exhibiting a tighter C:N stoichiometry in leaves, fine roots, and woody biomass. Discrepancies in leaf litterfall were possibly the result of under-reporting in the observations combined with the omission of the coniferous hemlock species in the model. Comparison of leaf mass and leaf litterfall stoichiometry suggests that leaf N resorption during leaf senescence was lower than the assumed 50 % in the model. In the observations, the mortality exceeded the wood production, which was not reproduced by the model. Instead, the model estimated that wood production and mortality were roughly balanced. The N flux associated with mortality was larger than the observed, probably because the observation-based estimate of mortality only accounted for the mortality of woody biomass, whereas the mortality term in the model also accounted for leaf, root, and fruit biomass.

The litter (also denoted as Oi) C pool was overestimated in the simulations compared with the observations. Given the discrepancy between observed litter C:N ratio and litterfall C:N ratio, it appeared likely that processes during litter decomposition already affected observed litter C:N, whereas the model's litter pool represents fresh, undecomposed litter. The simulations strongly overestimated soil C, as it was almost twice the observed value (Table ), where the fast-turnover pool contributed to 17 % of the total soil C pool and 18 % of the total soil N pool. The soil C:N ratio of the combined simulated fast and slow soil pools was similar to the observations. The observed soil C:N ratios declined strongly with soil depth, whereas, as assumed by the model, the simulated soil C:N ratio remained nearly constant across soil layers (Table S2 in the Supplement).

In the observations (Table S3 in the Supplement), the natural abundance δ15N values of leaves (-2.1 ‰) and coarse roots (-2.4 ‰) were most depleted and the wood (-1.2 ‰) was the most enriched vegetation pool. The simulated leaf δ15N values (-2.3 ‰ vs. -2.1 ‰) and wood (-1.2 ‰ vs. -1.1 ‰) were similar to the observations. The simulated δ15N in coarse roots (-1.4 ‰) was less depleted than in the observations (-2.4 ‰). The natural abundance δ15N levels in the litter layer and soil were more enriched than in the simulations, with the difference being largest in the deepest soil layers (Table S3).

In the experiment the tracer was sprayed on top of the litter layer. In the simulations, we added the incoming tracer to the N deposition flux, from which it directly entered the soil-solution NO3 pool of the uppermost soil layer. This slight difference may cause some discrepancies in the comparison, as in the model, the soil-solution NO3 pool of the first soil layer is directly accessible to plants and soil microbes. Given the timescale of this analysis, however, we assumed that delays in tracer transfer from the litter layer into the soil would not substantially affect tracer accessibility to plants and soil microbial biomass, and would therefore have little effect on our results.

Similar to observations, simulations showed a rapid δ15N enrichment of both leaves and fine roots in the year after the tracer addition, but the simulated decline of δ15N enrichment over the next decade was faster and larger than observed for these pools (Fig. ). Simulated leaf δ15N enrichment (Fig. a) showed a small increase already in 2007, but reached its highest level of 51.4 ‰ in 2008 (hereafter, all values reported are relative to the initial natural abundance δ15N), the year after tracer addition. One year later, the simulated δ15N enrichment decreased by more than 50 % and reached a low, steady enrichment (5.9 ‰) by 2012. The observations showed a smaller enrichment peak than the simulations (27.1 ‰), but the subsequent decline was slower and less pronounced than the simulated decline.

The fine root (Fig. b) observations showed a similar behavior to that of leaves, with a large increase in δ15N enrichment (46.6 ‰) during the year of labeling and some decrease in the first year after labeling (27.1 ‰), followed by a smooth decline by approximately 50 % over the following decade. The first observed δ15N value in 2007 includes only the uppermost layer (0–10 cm), whereas the other observations include fine roots in all the soil layers. The value in the upmost layer is larger than averaged across all soil layers, as it is where the largest amount of fine root biomass is located and the δ15N enrichment decreases with depth. The simulations approximately matched the observed δ15N enrichment peak (with a value of 25.5 ‰) observed in 2008, then showed a faster and larger decline than in the observations, dropping by two-thirds within five years and stabilizing at a lower level of δ15N enrichment compared to the observations.

The observed coarse-root δ15N values (Fig. c) showed a smaller peak in 2007 (33.1 ‰) than fine roots, but had a larger δ15N enrichment peak in 2008 than either the leaf or fine root pools. This suggests that the tracer was first captured by fine roots and then was subsequently transferred to the coarse roots in the following year in the observations. After peaking in 2008, the observed values declined, levelling off at approximately half of the δ15N peak by 2012. The simulations did not show such a distinct peak, only a small, gradual increase in the first four years (up to 6.6 ‰) after which the simulated δ15N remained nearly constant, decreasing only slightly to 6.1 ‰, well below the observed value (16.1 ‰).

The observed wood δ15N (including bark) increased by 4.9 ‰ in 2008 compared to the pre-labeling value and increased slightly further by 2012 (Fig. d). The simulations showed a larger δ15N enrichment peak (8.1 ‰), which declined to 2.7 ‰ within four years.

The observed soil δ15N enrichment after 2006 was much larger than the simulated one in all soil pools (Fig. ), which may partly result from the overestimation of the soil N pools in the simulation in both shallow (3.4-fold overestimated in simulation) and deep soil (2.2-fold overestimated in the simulation) (Fig. , Table S2). The observations showed a δ15N enrichment peak of 106.4 ‰ in the litter layer compared with the pre-experiment value and a decrease to about 10.4 ‰ over the following decade. The simulated 15N enrichment peaked later and only reached a value of 28.9 ‰ and a declined to a level of 4.6 ‰ over the following decade. In the observed shallow soil layer (0–10 cm), the tracer caused a δ15N enrichment of 9.2 ‰ in 2008 and a slow increase to a level of 12.4 ‰ in a decade. By contrast, the simulations showed an increase only of 3.2 ‰ and the δ15N enrichment started to slowly decrease, reaching 2.7 ‰ after a decade. The deeper soil layers experienced an increase in δ15N by 1.7 ‰ in 2008 and this value slowly increased up to 2.1 ‰ in a decade. The simulation, however, showed an increase of only 0.43 ‰ in a decade.

When only the soluble litter was considered (Fig. a) the agreement between simulation (simulated peak: 40.9 ‰) and observation was slightly better. The reason for this discrepancy lies in the immediate “adsorption” of the tracer signals when sprayed onto the litter surface, which cannot be represented by the model, since uptake of inorganic N by the litter pool is not a flux represented in the model. Instead, the δ15N tracer was added directly to the soluble NO3 pool, from which it was available for plant and microbial uptake, the latter of which generates fast SOM, as well as leaching and gaseous emissions. In the simulation, the δ15N signal in the litter originated solely from litterfall of enriched plant material, and thus most of the added δ15N tracer was immobilised by plants or soil microbes. In fact, while Fig. b and c show that the model did not capture the magnitude of the bulk SOM δ15N signal, when only the fast SOM pool was considered, the observed change in SOM was well captured in the topsoil layer (peak enrichment of 12.2 ‰) (Fig. b). The observed δ15N values in the deeper layers showed little variation, similar to the simulated changes, although a slight propagation of the surface signal to lower layers can be seen in later years.

The vertical distribution of the 15N tracer showed that the observed soil was more enriched in 2006 than the simulated soil throughout the depth profile (Fig. ). Furthermore, the observed δ15N signal in the soil was most depleted in the top layer, and most enriched at around 0.25 m depth. In the simulations the soil became steadily more enriched towards deeper layers in 2006.

The temporal changes in soil δ15N were much larger in the observations than in the simulation, as already shown in Fig. . In 2008 the increase in observed soil δ15N extended to a depth of 30 cm, and during the following years the enrichment continued to increase slowly in the upper soil layers. In the simulation, the temporal dynamics were different. The top layer was most enriched in 2008 and subsequently became depleted. In 2008, the signal had not yet reached the layers below 20 cm in the simulation. During the following years, the enrichment resulting from the δ15N tracer addition gradually propagated to deeper soil layers. Comparison of the observed and simulated soil N profiles showed that the simulation overestimated the soil N pool in the top layer, whereas below this layer the vertical gradient was more stable and similar to that observed (Fig. ).

The total recovery of the tracer was close to 100 % in the simulation (Fig. ), whereas it was around 70 % in the observations. Both the observations and the simulation showed a large fraction of the δ15N signal remaining in the shallow soil layer, and the relative contribution of deeper soil layers increased only slowly over time. The observations showed a noticeable fraction of the δ15N signal remaining in the litter layer, which was almost depleted by 2017, whereas in the simulation the litter pool was of minor importance.

The δ15N signal in coarse roots was pronounced in the observations, but not in the simulation. Both the observations and the simulation showed a contribution from the fine roots. The simulation showed a substantial contribution from the wood pool already in 2008, which remained substantial thereafter. In the observations the transfer of δ15N to the wood pool occurred in 2008 and increased in 2012 (wood was not measured in 2017; values for 2012 were retained for visualization). This difference occurred partly because we attributed all labile and reserve 15N to the woody pool, whereas in reality the signal may be partitioned between coarse roots and woody reserve storage. The simulation showed a large fraction of the recovered δ15N in leaves in 2008 (9.5 %), while the observations showed a recovery of only 1.9 % for that year. When soil and plant pools were considered separately, the observations showed a prominent signal in the soil already in 2008, and its relative contribution decreased over the following years (Table ). Recovery of the label in the plant pools peaked at 11 % the year after the tracer addition, then stabilized at 8 % in years 5 and 10. In the simulations, the relative contribution to plants was larger than in the observations in 2008, but was similar to the observations in 2012 and 2017 (Table ).

In order to track how much of the δ15N signal remained in the ecosystem over longer timescales, we made a scenario simulation for an additional 30 years. The simulated total recovery rate of the system remained high until the end of the simulation, suggesting a relatively closed simulated nitrogen cycle at the site (Fig. b). Most of the signal remained in the soil, with the deeper soil layers gaining δ15N signal from the top soil layers. Over time, the woody pool became an increasingly important vegetation pool for storing the δ15N signal.

The vertical profile of simulated δ15N in soil (Fig. b) became more uniform over time, while the uppermost soil layers became depleted. A comparison of the simulated δ15N values in different ecosystem pools in 2017 and 2048 (Table ) indicates that all vegetation pools, except the wood pool, showed a gradual decrease in δ15N during this period. The δ15N in the litter was slowly decreasing. In the soil, δ15N decreased in the fast SOM pool but increased in the slow SOM pool.

Our simulation showed how the 15N tracer signal applied to the forest floor was transferred among ecosystem pools over decadal timescales. Despite some conceptual differences in the description of the manipulation experiment, the QUINCY model successfully simulated many features observed in the observations. The comparison with the observations was especially interesting when evaluating the N turnover rates of the vegetation pools. These suggest that the model's turnover rates of N were too fast for the leaves and fine roots. This occurred despite the seemingly higher N retention of leaves during senescence in the model. It is important to note that the model underestimated the C and N pool size of fine roots and C pool size of the leaves, and this might have some influence on the simulated dynamics. Although the simulated total plant C pool (12 771 gCm-2) was close to the observed value (11 140 gCm-2), the other vegetation C pools, except for wood, were underestimated. This suggests that QUINCY has some difficulties in representing C allocation. The simulated total N in vegetation (86.3 gNm-2) was over double the observed amount (39.2 gNm-2), reflecting the overestimation of soil N.

Since this is a deciduous forest and the turnover time of fine roots is set to 0.7 years in the model, it is likely that the 15N tracer was transferred too quickly away from the compartments that are used to build leaves and fine roots. In the observations, the coarse roots showed a large 15N peak, whereas the simulation showed only a gradual increase. The current model structure does not accommodate a quick response from the coarse roots, that currently have a turnover time of eight years. The observations showed a larger signal in the coarse roots, because the movement of the signal to the upper tree parts required transport, e.g., via diffusion or xylem transport, and additionally the coarse-root observations included xylem tissue . In QUINCY, we could account for the larger xylem concentration in coarse roots by partitioning the labile and reserve pools among coarse roots, leaves and woody biomass, whereas in this study for simplicity we have assumed they were allocated to sapwood. Given the simulated peak 15N signal in labile+reserve pool, partitioning 11 % of this pool to coarse roots would reduce this mismatch. This suggests that the timescale of the transfer of the 15N signal within the plant (from coarse roots to woody and leaf compartments), is potentially significantly larger (on the order of months to few years), than the nearly instantaneous redistribution of nitrogen in the plant as assumed by the model.

Finally, it should be mentioned that the definition of coarse and fine roots differed slightly between the model and the observations. In the observations, roots thicker than 1 mm were considered coarse roots , whereas the threshold was 2 mm in the model. The 1 mm cut-off point for the fine roots in these observations was chosen to align with earlier studies conducted at this site . It has also been shown that fine roots with diameters less than 1 mm constitute the majority of fine-root biomass and exhibit the greatest sensitivity to nutrient availability in this and similar forest types . We therefore do not expect this discrepancy between the simulation and observations to have a significant impact on the comparison.

In the model, the tracer was added to the top layer of solute NO3 which enabled direct access by plants and bypassed the litter layer. The trees took up this nitrogen into labile and reserve plant N pools in the model. This caused the simulated plants to have more of the 15N than the soil than observed at the end of 2007. However, by 2008, the observations showed a relatively larger fraction of the tracer in the plants than the simulations. This difference suggests that the simulated plant pools had too high turnover rates for nitrogen, as was also seen separately for the leaves and fine-root compartments. In addition, the strong uptake of the tracer by the litter layer and followed by the gradual release of 15N could also have influenced the plant signal in the observations. In the observations, the recovery of the 15N in coarse roots remained substantial during the last years of the experiment and this supports the hypothesis that the tracer was still being slowly released from the litter and being transported to the plants. Generally the relative recovery of 15N between plants and soil was comparable between the simulation and the observations in later years. However, during the few first years the model estimated nearly twice the recovery of 15N in plants (∼21 %) as the observations (∼11 %), in line with what has been found out in the earlier modeling studies that modeled general N dynamics rather than explicit 15N cycling e.g..

Implementation of nitrate assimilation by soil microorganisms was a key process for successfully simulating the tracer experiment at the Arnot Forest. The observed and measured ecosystem N pool magnitudes differed, which also influences the interpretation of 15N temporal dynamics. These differences were pronounced in the soil, with the modeled total soil N pool (1656 gNm-2) was more than twice as large as the observed pool (744 gNm-2). Since in the model the slow soil pool was large, but it was not taking up a lot of the 15N signal due to its lower turnover rate, its large size caused low values of the 15N tracer signal. This effect was demonstrated by isolating the fast pools (Fig. ), that better captured the observed changes in the magnitude of the 15N signal in the uppermost soil organic matter layer.

In the observations, the 15N signal reached deeper soil layers, down to 0.35 m (observed layer depth 0.3–0.4 m), after only one year and at 0.45 m (observed layer depth 0.4–0.5 m) within five years. In the simulation, the downward movement of the 15N signal was slower. Several factors may have contributed to this finding, including a too large uptake of the signal by vegetation and soil microbes, a too small movement of NO3, or an underestimation of the transport by bioturbation. Bioturbation in the model is represented as a simple diffusive flux as in , but with the rate declining with soil depth driven by a decline on the fine root biomass density. This decline might be too strong in QUINCY, but also the amount of C allocation to roots was underestimated in the model, which might have contributed to a too weak bioturbation flux. A further contribution may be that a fraction of the leaching flux was simulated to occur via lateral flow and not deep drainage, effectively partitioning the leaching loss into a lateral component that prevented a fast and deep penetration of the 15N signal in the soil.

In the observations, the top soil layers became enriched as the 15N from the overlying litter layer was incorporated into the soil organic matter. The top layers showed some depletion by 2018. The simulation showed larger changes in the top layers over time and also the 15N tracer was also transported to deeper soil layers, which resulted in a flattening of the vertical 15N soil profile (Fig. ). These differences between the model results and observations stem from conceptual differences. It is likely that in the observations the 15N tracer got attached to the litter layer and was being released slowly still after several years, resulting in a continued enrichment of the top soil layer. This is possible in the real world because the observed litter pool includes early stages of decomposition, typically associated with an increase in N concentration due to immobilisation of mineral N by decomposers, which facilitates the rapid incorporation of 15N. In the model, the litter pool only contains fresh litter and any derivates, including immobilized mineral N, that are transferred to the fast SOM pool. Disentangling these conceptual differences from any model structural deficiencies is challenging. It is of note that the considerable overestimation of the soil N pool that occurred across all soil depths partly contributed to the simulated soil 15N signal being more depleted than observed.

The 15N tracer was applied as NO3 remained for a very long time in the soil according to both the observations and the model analysis. This is in line with other natural forest ecosystem 15N studies , where the largest compartment of recovered 15N was also the soil organic matter. It has been suggested that the NO3 can be incorporated into soil organic matter through abiotic processes . These kinds of processes are not yet implemented in the current model set-up.

The natural-abundance soil δ15N differed between the observations and the simulation in 2006, prior to the start of the experiment, with observed soil δ15N values being more enriched than those in the simulation. Greater soil δ15N enrichment is often associated with increased N losses via gaseous pathways which strongly discriminate against δ15N . This could indicate a less closed N cycle and weaker N limitation in the observations than in the simulation. However, following the initial tracer loss, the observed δ15N recovery remained relatively stable over time, suggesting relatively efficient long-term N retention and a relatively closed N cycle. The mycorrhizal fungi present at the site are known to strongly fractionate N isotopes , which may have also contributed to the enriched soil δ15N values in the observations. Leaf stoichiometry can be considered a proxy for leaf N status . The simulation showed a lower leaf C:N ratio than the observations, suggesting weaker N limitation in the simulation.

N deposition declined during the study period. In N-limited ecosystems such a decline might lead to a more closed N cycle. However, in boreal forests, the role of increasing atmospheric CO2 has been shown to be a more important driver for a reduction in N availability than changes in N deposition and a similar response may occur in temperate forests. Overall, changes in N limitation during the 11-year study period are unlikely to have influenced our conclusions.

The ecosystem-scale recovery of the 15N tracer is very challenging . Recovery of the tracer remained close to 100 % in the model throughout the decade following tracer addition. In the observations, recovery was also close to 100 % in the days following each tracer addition, but declined to 77 % by the end of the year, largely due to losses during midsummer . This N was most likely lost via gaseous pathways (denitrification), as measured leaching losses of 15N during the first seven months were very low (<1 %). This recovery is very similar to the average recovery rate of 77.9 % for deciduous forests .

The large difference in total recovery rates between the observations and the simulation originates from the way the 15N tracer input to the ecosystem is represented in the model. In our simulation the 15N tracer entered directly into the soluble NO3 pool of the uppermost soil layer from where it can be taken up by plants or soil microbes, before any gaseous losses occurred. Some gas losses are likely to occur, but because their magnitude cannot be prescribed with confidence, it was not possible to explicitly represent them in the model with this approach. Despite the discrepancy in the total recovery rates, we can use our results to investigate how the 15N propagates through the different ecosystem compartments.

Because of the way the tracer experiment was represented in the model, the litter layer also did not show the pronounced recovery of the 15N signal observed in the measurements (Fig. ). Nevertheless, the model captured several key characteristics of the observations: a large recovery of the signal in the top soil layers and an increasing recovery of the signal with time in the lower soil layers. The simulation supported a long-term retention of N within the ecosystem. This was particularly evident in the scenario results for 2048, as the total recovery rate remained high (88 %). The movement of the 15N tracer continued into deeper soil layers and the wood became an increasingly important storage pool of the 15N tracer over time, with only very low losses of the 15N tracer from the ecosystem. Approximately 11 % of the recovered tracer resided in the plant organs, of which more than 80 % was stored in wood (corresponding to 9 % of total recovery). Due to the high C:N ratio of wood, wood can potentially sequester more carbon per unit of nitrogen than other plant compartments. For NO3 tracer experiments in temperate forests, found that almost half of the tracer in plants would reside in stem and branches, that would correspond to wood in our simulations, and more than half in roots and leaves. Our model simulations suggested a larger fraction of the plant compartment's tracer residing in wood than that reported by , implying a greater potential for carbon sequestration. In addition, our simulations cover a longer time span than is currently available from observations.

This study has a number of uncertainties. In the measurements the heterogeneity and large mass of the soil pose a considerable challenge . In the simulation we can capture the overall behaviour based on process understanding. To determine whether the issues identified at Arnot Forest are generic model deficiencies, it would be valuable to extend the analysis to additional experiments.

Since our aim was to assess the N retention at the ecosystem scale and 15N recovery in the different ecosystem compartments, we consider the model results to be sufficiently robust for this type of analysis. used Latin hypercube sampling to assess the influence of uncertain fractionation rates for different processes e.g. in the QUINCY model. They found that uncertainties of these fractionation rates did not affect the overall conclusions of their study.

While the QUINCY model covers most fluxes of the terrestrial N cycle, a few processes are missing: it has long been known that some plants are able to also acquire organic nitrogen directly from soil organic matter , and that symbioses with ectomycorrhizae can also facilitate this uptake , but the QUINCY model currently does not include these processes. Their omission might be relevant because half of the trees at the site had ectomycorrhizal associations.

The simulation in this study used the basic soil model included in QUINCY. To include abovementioned processes it would be preferable to use as a starting point the more detailed soil model of QUINCY, Jena Soil Model (JSM) , which explicitly represents microbial processes and organic matter stabilisation. This new model has been shown to improve the simulation of soil profiles at different sites owing to its representation of microbial process and its inclusion of organo-mineral association. Including more explicit process representations instead of relying on first-order kinetics brings the simulation results closer to reality and simulating this tracer study would likely also benefit from using JSM. Using JSM might also help in more realistic representation of immobilization, that has often been identified as a limitation in N cycle modelling studies . To fully benefit from the increased process representation of the JSM in future studies, observations of microbial biomass, mineral-associated organic carbon and their 15N values as well as dissolved organic carbon would be valuable for improving our understanding of the underlying processes.

15N tracer studies have also shown that the 15N signal is incorporated into old wood . This process is not described in the model and would not be captured. In reality, the canopy can also directly take up atmospheric N deposition in some cases (but see ) and thereby influence the carbon uptake of vegetation , even though the soil is the most important pathway for the N entering the plants . In the Arnot Forest experiment we simulated the 15N tracer experiment, in which the tracer was sprayed onto the forest floor. The application of the tracer on the forest floor results in lower recovery rates of the tracer than would happen if it was applied on the canopy . At present we do not have a mechanism in the QUINCY model that would include a mechanism for the uptake of N deposition via leaves.