The empirical data are from Wardlow Hay Cop (henceforth referred to as Wardlow), a long-term experimental grassland site in the Peak District National Park (UK) (Morecroft et al., 1994). Details of empirical data collection are available in Supplement Sect. S1. There are two distinct grassland communities occurring in close proximity: acidic (National vegetation classification U4e) and limestone (NVC CG2d) semi-natural grasslands (Table S2). Both grasslands share a carboniferous limestone hill, but the limestone grassland sits atop a thin humic ranker (Horswill et al., 2008) and occurs predominantly on the hill brow. In contrast, the acidic grassland occurs in the trough of the hill, allowing the accumulation of windblown loess and the formation of a deeper soil profile of a palaeo-argillic brown earth (Horswill et al., 2008).

Despite contrasting soil types, both the acidic and limestone grasslands are largely P-limited (Morecroft et al., 1994; Carroll et al., 2003), though occasional N and P co-limitation can occur (Phoenix et al., 2003), and more recently, positive growth responses in solely N-treated plots have been observed, in line with the latest understanding that long-term N loading may increase P supply by increasing phosphatase enzyme activity (Johnson et al. 1999; Phoenix et al.2004; Chen et al. 2020).

Nutrients (N and P) have been experimentally added to investigate the effects of elevated N deposition and the influence of P limitation (Morecroft et al., 1994). Nitrogen treatments simulate additional N deposition to the background level, and the P treatment acts to alleviate P limitation. Nutrients are added as solutions of distilled water and applied as fine spray by backpack sprayer and have been applied monthly since 1995, and since 2017 bi-monthly. Nutrient additions are in the form of NH4NO3 for nitrogen and NaH2PO4 for phosphorus. Nitrogen is applied at rates of 0 (distilled water control – 0N), 3.5 (low nitrogen – LN), and 14 g N m-2 yr-1 (high nitrogen – HN). The P treatment is applied at a rate of 3.5 g P m-2 yr-1 (phosphorus – P).

Data collected from the Wardlow grasslands for the purpose of this work are aboveground biomass C, SOC, and total N, which is assumed to be equivalent to modelled SON. These new data are combined with total P data that were collected by Horswill et al. (2008) at the site (Horswill et al., 2008). Summaries of these data are available within the Supplement (Table S1), and details of their collection and conversion to model-compatible units are in Supplement Sect. S1.

We use data from the Wardlow limestone and acidic grasslands to explore the potential role organic P access may have in determining grassland nutrient limitation when exposed to long-term N deposition and more recently experimental nutrient manipulation. We use environmental input data collated from Wardlow to drive model processes. Empirical data regarding contemporary soil C, N, and P for the contrasting grasslands are used to calibrate the initial size of the weatherable P pool within the model and to allow access to organic cleaved P to vary to account for patterns in the data. We do not aim to perfectly replicate the Wardlow grasslands but rather use the unique opportunity that Wardlow provides to test our understanding of such P-limited ecosystems and how our conceptualisation of P-access mechanisms within the model may affect them. In addition, we can use the model-simulated grasslands to investigate the potential effects of long-term N deposition and nutrient manipulation on ecosystems which may differ in their relative availability of different P forms.

The model calibration selected parameter values for PWeath0 and PCleaveMax that indicate contrasting use of P sources by the two simulated grasslands, with the acidic grassland capable of acquiring more P from organic sources having a PCleaveMax value of 0.32 g m-2 per season compared to the limestone, with a value 10 times smaller at 0.03 g m-2 per season. Conversely, inorganic P availability was greater in the limestone grassland due to the larger weatherable pool of P, PWeath0, at 300 g m-2 compared to 150 g m-2 in the acidic grassland.

The selected parameter combinations resulted in the model simulating the acidic grassland as N-limited and the limestone as P-limited, with reasonable congruence between observed and modelled data. The outputs for the calibrated model are shown in Fig. 2 against the observations for above-ground biomass C, soil organic C, N, and total phosphorous (TP) for both the acidic and limestone grasslands (Fig. 2). Raw data used for Fig. 2 are provided in Supplement Tables S4 and S5.

Overall, N14CP more accurately simulated the magnitude of limestone grassland C, N, and P pools than the acidic grassland, and it generally captured the pattern of responses to nutrient treatment, albeit this is not always supported by high r2 values. The model estimates of above-ground biomass C are broadly aligned with the observations: capturing variation between the grasslands and treatments (r2=0.58) and on average overestimating the magnitude by 12.9 % (SE ± 11.9) and 12.1 % (SE ± 9.4) for the acidic and limestone grasslands respectively (Fig. 2a).

Soil organic C on average was slightly overestimated (7.1 % with SE ± 3.3) for the limestone grassland (Fig. 2b), with a larger average overestimate for the acidic grassland (39.9 % with SE ± 6.8). However, in this latter case the variation between treatments was better captured. Despite a low r2 value for SOC (0.01), the model broadly captured the patterns we observe in the empirical data, with N addition increasing SOC in the acidic grassland and P addition increasing SOC in the limestone grassland. However, the intermediate increase in SOC with P in the acidic grassland is not captured by the model, nor is the magnitude of the negative effect of LN treatment on limestone SOC.

Simulated magnitudes of SON are well-aligned with observations for the acidic grassland, with an average error of 2.3 % (SE ± 3.2), whilst SON for the limestone grassland was on average underestimated by 17.8 % (SE ± 3.6) (Fig. 2c). The variation between treatments was better captured for acidic than limestone SON but was overall reasonable (r2 = 0.39).

Finally, the model overestimated TP (defined in the model as organic P plus sorbed P) by an average of 6.0 % (SE ± 4.3) for the limestone but underestimated by 54.7 % (SE ± 8.0) in the acidic grassland, which was the least accurately predicted variable out of those investigated (Fig. 2d). With only two empirical data points for TP across only two nutrient treatments, it is difficult to discern the relationship between treatments and TP, so an r2 value is of little relevance here.

Modelled acid grassland NPP remained N-limited from 1800 through to 2020 under most nutrient treatments (Fig. 3). Nitrogen deposition increased the potential NPP through time, and the grassland moved toward co-limitation in the LN treatment (i.e. the N and P lines were closer) but remained N-limited (Fig. 3b). In the HN treatment, the acidic grassland shifted to P limitation as N-limited NPP surpasses P-limited NPP (Fig. 3c).

The simulated limestone grassland was also initially N-limited but was driven through a prolonged (ca. 100 years) state of apparent co-limitation until clearly reaching P limitation in 1950, solely as a result of N deposition (Fig. 3). In the 0N treatment, the grassland remained P-limited, but the potential NPP values for N and P are similar, suggesting the grassland is close to co-limitation (Fig. 3e). The LN and HN treatment amplified pre-existing P limitation, lowering the potential NPP of the grasslands (Fig. 3f, g). With the addition of P in 1995, P limitation is alleviated, and the ecosystem transitions to a more productive N-limited grassland (Fig. 3h).

Another way to interpret the extent of nutrient limitation within N14CP with specific reference to P demand is to assess the rate of P cleaving through time. These data corroborate the N and P-limited NPP data, showing that in the limestone grassland, the maximum amount of cleavable P is accessed by plants in the 0N, LN, and HN treatments from approximately 1900 through to the end of the experimental period in 2020 (Fig. S1, Table S13), highlighting its consistent state of P limitation.

Conversely, while cleaved P is used in the 0N treatment in the acidic grassland, it occurs at approximately one-third of the total rate; hence the grassland is not entirely P-limited (Fig. S1, Table S9). The LN treatment increases the rate of access to cleaved P, and HN causes it to reach its maximum value, confirming the shift to P limitation suggested by the NPP data (Fig. S1, Table S9). Soil organic P cleaving does not occur in the P-treated plots of either grassland.

The model allows the temporal trends and responses to nutrient additions to be further explored. Figure 4 provides the temporal responses for the treatments and Fig. 5 a full nutrient budget for the year 2020. Full data for changes in soil C, N, and P and plant biomass C pools since the onset of large-scale N deposition (1800 within the model) for both grasslands are included in Supplement Table S14. All data used for determining responses of biomass C and soil organic C, N, and P pools to experimental nutrient additions are in Supplement Tables S15 (acidic) and S16 (limestone).

This is the first instance in which N14CP, and to the best of our knowledge any other integrated C–N–P cycle model, has explicitly modelled P-limited ecosystems and investigated their responses to N deposition and additional nutrient treatments. By using empirical data from long-term experimental grasslands to drive and calibrate N14CP, we could test the model's ability to simulate two contrasting P-limited grasslands and how organic P access may affect this ability. While the purpose of this work was not to explicitly reproduce the Wardlow grasslands within N14CP, by comparing data from Wardlow to the simulated grasslands, we can simultaneously develop our understanding of the model's representation of under-studied P cycling processes and contextualise what this may mean for empirical systems such as Wardlow.

The model suggests that the acidic grassland was characterised by high access to organic P, with comparatively low inorganic P availability, whereas the limestone grassland was the opposite, with low organic and high inorganic P availability. These simulated differences could reflect the relative availability of different P sources at Wardlow. As the acidic grassland formed in a hillside depression, loess has accumulated, thickening the soil profile and distancing the plant community from the limestone bedrock. The plant rooting zone of the acidic grassland is therefore not in contact with the bedrock, and roots almost exclusively occur in the presence of organic P sources which can be cleaved and utilised by plants (Caldwell, 2005; Margalef et al., 2017). Conversely, the limestone grassland soil rarely exceeds 10 cm depth, and the rooting zone extends to the limestone beneath, providing plants with greater access to weatherable calcium phosphate (Smits et al., 2012).

Such parameter combinations allowed for reasonable congruence between empirical and simulated data, with an average discrepancy of only 6.6 % (SE ± 9.1) and 1.2 % (SE ± 4.4) for the acidic and limestone grasslands respectively across all variables (Table S5). However, model performance differed greatly between the two grasslands. For instance, the model accurately captured the magnitude of limestone C, N, and P data and their expected P-limited responses to nutrient treatment but was less effective at simulating the acidic grassland. N14CP did not simulate an increase in biomass C or SOC with P addition in the acidic grassland, instead simulating a solely N-limited grassland. While this may be expected of a model that employs a law-of-the-minimum approach, N14CP has a number of mechanisms to account for N and P interdependence, meaning that in principle, it is capable of simulating positive responses to LN, HN, and P treatment, as observed in the empirical data from 2019 (Sect. 2.2.2).

The overestimation of acidic C pools and underestimation of total P suggest that the model is simulating that too much organic P is being accessed by plants in response to N addition and transferred into plant biomass pools (Fig. 2d). Few parameter sets were simultaneously able to simulate the magnitude of the empirical TP pool and the positive response of biomass to N addition in the acidic grassland. This may also be due to limitations in the empirical P data, as P data used for calibrating P cycling were available for only two nutrient treatments and represented total soil P, not organic P. While we acknowledge the technical and theoretical issues associated with distinguishing between organic and inorganic P pools (Lajtha et al., 1999; Barrow et al., 2020), such distinctions would help in understanding this discrepancy and likely improve the model's ability to simulate P-limited systems, particularly when organic P availability may be important.

Additionally, N14CP's representation of organic P cleaving likely underestimates the ability of soil to rapidly occlude and protect organic P that enters solution. For example, inositol phosphate, a major constituent of organic P, has been found to be used extensively by plants grown in sand but is hardly accessed by plants grown in soil (Adams and Pate, 1992). Such organic phosphates become strongly bound to oxides in the soil, protecting them from attack by phosphatase enzymes (Barrow, 2020). This may be particularly prevalent in the acidic grassland at Wardlow where N deposition has resulted in acidification and base cation depletion (Horswill et al., 2008), potentially enhancing the formation of iron and aluminium complexes and immobilising P (Kooijman et al., 1998).

In addition to physico-chemical processes reducing P availability, in P-limited grassland soils, microbial processes may be dominant drivers of ecosystem P fluxes (Bünemann et al., 2012). For instance, while mineralisation of organic P may increase inorganic P in solution (Schneider et al., 2017), this can be rapidly and almost completely immobilised by microbes, particularly when soil P availability is low (Bünemann et al., 2012). As the model lacks a mechanism for increasing access to secondary mineral P forms comparable to organic P cleaving, and microbial P immobilisation is incompletely represented for P-limited conditions, it is possible that the uptake of organic P by the acidic grassland in the model is exaggerated.

The model's inability to simulate a positive response to both N and P addition in the acidic grassland may be an unintended consequence of the downregulation of N fixation by N deposition included within N14CP (Davies et al., 2016b). While this representation is appropriate (Gundale et al., 2013), when N deposition exceeds fixation (as at Wardlow), fixation is essentially nullified (as in Tables S7, S11), meaning deposition becomes the sole source of N to the grassland. This in effect removes the dependence of N acquisition on P availability and could make modelling behaviour akin to N–P co-limitation (Harpole et al., 2011) under high levels of N deposition challenging. This suggests that current C–N–P cycle models that employ Liebig's law of the minimum can provide a broad representation of multiple variables by calibrating access to both organic and inorganic P sources (Davies et al., 2016b), provided the ecosystem in question's limiting nutrient leans towards N or P limitation. Furthermore, where access to organic P forms is likely to be lower, as in the limestone grassland, model performance may improve. This could be further explored by allowing N fixation limits in the model to adapt to P nutrient conditions or by attenuating the suppression of N deposition on N fixation, to represent acclimatisation of N fixers to greater N availability (Zheng et al., 2018).

Ultimately, differences in modelled accessibility to organic forms of P enabled N14CP to distinguish between the two empirical grasslands and simulate the magnitude and pattern of data with reasonable accuracy, albeit with the previously mentioned caveats.

While the model's estimation of PCleaveMax for the acidic grassland is likely overestimated, the model experiment has highlighted that differences in organic versus inorganic P availability are a key determinant of an ecosystem's nutrient limitation and consequently how they respond to changes in anthropogenic N and P availability. For instance, while being exposed to the same background level of N deposition and the same magnitude of experimental treatment, the modelled acidic grassland was able to stimulate growth in response to LN and HN treatment, whereas the modelled limestone grassland was negatively affected by it.

Nitrogen addition increases plant demand for P and can shift ecosystems toward a state of P limitation or increase the severity of limitation where it already exists (Menge and Field, 2007; An et al., 2011; Goll et al., 2012). Consistent with this, both simulated grasslands saw SOP decline with LN and HN treatment, worsening P limitation in the limestone grassland and depleting the SOP pool in the acidic grassland. As P cleaved from organic pools is the least bioavailable within the model hierarchy (Sect. 2.2.3), this is indicative of increasing P stress in both grasslands. While SOP declined in both grasslands, the responses of available and biomass P to nutrient treatments differed markedly between the grasslands. Due to the higher rate of PCleaveMax in the acidic grassland, more P was in plant-available forms, and hence P does not become the limiting factor under N treatments (Table S8). Conversely, available and biomass P decline under LN and HN addition in the limestone grassland (Table S12), highlighting how the grassland's PCleaveMax capability is insufficient to meet increased P demand.

Such high access to organic P sources in the modelled acidic grassland likely led it to respond to nutrient enrichment in an N-limited manner, increasing productivity in response to N deposition and LN and HN treatments as the model's limiting nutrient stimulated plant growth. Detrital C inputs from plant biomass are the primary source of SOC accumulation within N14CP (Davies et al., 2016b), and as such, changes in SOC integrate long-term trends in net primary productivity in systems where external nutrients are supplied. The provision of additional N in the modelled LN and HN treatments therefore led to large increases in biomass accumulation and consequently almost linearly increased SOC (Fig. 4c).

Similar increases in N-limited grassland SOC under N addition have been shown, resulting from significant increases in below-ground carbon input from litter, roots (He et al., 2013), and detrital inputs (Fornara et al., 2013), mechanisms similar to those reported by the model. Similarly, Tipping et al. (2017) used N14CP to show that N deposition onto N-limited UK ecosystems ubiquitously increased SOC storage by an average of 1.2 kg C m-2 (ca. 10 %) between 1750 and 2010 (Tipping et al., 2017).

Despite its P-limited condition under the HN treatment (Fig. 3c), the acidic grassland continued to accumulate biomass with N addition as the grassland's greater access to topsoil SOP (Table S8) allowed it to acquire sufficient P to stimulate additional growth but not necessarily to alleviate P limitation. This is consistent with the acidic grassland at Wardlow, where N treatment stimulated root surface phosphatases, likely supplying more SOP to plants (Johnson et al., 1999). Our simulated acidic grassland therefore supports the hypothesis that prolonged N deposition may increase SOP access to such an extent that P limitation is alleviated and growth can be stimulated (Chen et al., 2020). Organic P release from SOM and its potential immobilisation are poorly represented in models, and we encourage further study aimed at quantifying these processes (Chen et al., 2020; Janes-Bassett et al., 2020; Phoenix et al., 2020). However, such high rates of SOP access only occurred under experimental LN and HN treatments, and in reality, such rapid degradation of SOP may eventually degrade the pool to such an extent that P limitation soon returns.

Conversely, biomass C and SOC in the modelled limestone grassland responded positively to P addition, via similar mechanisms to the N response in the modelled acidic grassland. However, in contrast to the acidic grassland, N addition caused declines in limestone biomass and SOC, the former of which has been observed at the limestone grassland at Wardlow (Carroll et al., 2003). Reductions in limestone biomass C (and consequently SOC) in the model are a combined result of reductions in bioavailable P (Table S12), occurring via N-driven increases in stoichiometric P demand, in addition to an inability to access sufficient P from the SOP pool (Table S14). Plants therefore cannot meet P demand, and new biomass is insufficient to replace senesced plant material, decreasing net biomass C input to the SOC pool. This suggests that in P-limited limestone grasslands such as at Wardlow, where access to organic P forms may be comparatively limited, N deposition may worsen pre-existing P limitation and reduce ecosystem C stocks (Goll et al., 2012; Li et al., 2018).

While N14CP is a fairly simple ecosystem model by design, it is one of few models to integrate the C, N, and P cycles for semi-natural ecosystems and has been extensively tested against empirical NPP and soil C, N, and P data (Davies et al., 2016a, b; Tipping et al., 2017, 2019; Janes-Bassett et al., 2020). Previous work with N14CP has identified the need to enhance its ability to simulate organic P cycling (Janes-Bassett et al., 2020), which we aimed to do in this study by using long-term experimental data from contrasting P-limited grasslands.

N14CP's simplified representation of plant nutrient pools and plant control over nutrient uptake is largely controlled by stoichiometric demand (Davies et al., 2016a) and does not incorporate many plant strategies for P acquisition (Vance et al., 2003). Indeed, by allowing PCleaveMax to vary to account for empirical data, we attempt to somewhat increase plant control over organic P uptake. We acknowledged earlier that such an approach likely underestimates the ability of soil surfaces and microbes to protect newly cleaved P from plant uptake. As such, where we may expect access to organic P to be high, such as the acidic grassland at Wardlow, such a modelled representation of plant-mediated P access may lead to unrealistic depletions in soil P and increases in biomass and soil C, and we would encourage further work aimed at improving model representation of plant controls on organic P cycling (Fleischer et al., 2019).

While we feel incorporating a suite of plant strategies for acquiring P would represent over-parameterisation, we acknowledge that a modelled equivalent to PCleaveMax for accessing inorganic P forms is lacking, such as carbon-based acid exudation to increase mineral P weathering (Achat et al., 2016; Phoenix et al., 2020), which likely contributes toward the poor representation of the acidic total P pool. Biota-enhanced P weathering and nutrient redistribution by mycorrhizal hyphae are important for nutrient cycling (Quirk et al., 2012), and fungal community structure and function are strongly influenced by perturbations in the C and N cycles (Moore et al., 2020). Such processes are not included within N14CP as the extent to which weathering can be controlled by such mechanisms and the manner in which these can be represented in C–N–P cycle models are debated (Davies et al., 2016b).

Currently, N14CP assumes C to be in unlimited supply, with its uptake by plants and consequent input into soil pools controlled by C : N : P stoichiometry; hence C availability has little effect on N and P dynamics within the model. Increasing atmospheric CO2 may increase nutrient availability, as plants may reallocate additional carbon resources toward nutrient acquisition (Keane et al., 2020), or elevated CO2 (eCO2) may increase limitation of other nutrients such as N (Luo et al., 2004). The inclusion of eCO2 into N14CP poses a particularly enticing research opportunity, and we aim to use this study as a foundation for future work to include this process.