The study was performed inside pasture paddock no. 8 of the Grünschwaige Grassland Research Station near Freising, Germany (Schnyder et al., 2006). Mean annual air temperature from 2006 to 2012 was 9.3 ∘C, and mean annual precipitation was 743 mm, as measured at the Munich airport meteorological station 3 km from the field site. The soil is a Mollic Fluvisol, with a shallow topsoil of low water holding capacity (66 mm plant-available field capacity) overlying coarse calcareous gravel. The depth to the groundwater table is around 1.5 m.

During the main vegetation period (mid-April to beginning of November) the paddock was grazed continuously by Limousin suckler cows (Schnyder et al., 2006). Animal stocking density was adjusted periodically to maintain a constant sward height of about 7 cm. This management system aimed at maintaining a constant sward state by continuously balancing pasture grass production and consumption by the grazing cattle.

Precipitation water was collected following events during the vegetation periods of 2007 to 2012 and during winter 2007/2008 (see Methods S1 in the Supplement). Leaf, stem, soil, groundwater and atmospheric moisture samples were collected on non-rainy days, between 11:00 and 16:00 CEST (Central European Summer Time). Sampling occurred at approximately fortnightly intervals during the vegetation periods from April 2006 to September 2012. Samples were collected at random locations in an area of about 1 ha in the vicinity of an eddy flux tower installed near the centre of the paddock. On each date, two replicate samples of leaf, (pseudo-)stem and soil were collected. Soil samples were taken at two depths (7 and 20 cm) using an auger. Leaf and stem samples were obtained as mixed-species collections of the co-dominant species: four C3 grasses (Lolium perenne, Poa pratensis, Phleum pratense, Dactylis glomerata), one rosette dicot (Taraxacum officinale) and one legume (Trifolium repens). Each leaf sample included all leaf blades, including the exposed part of the growing leaf but excluding senescing leaves (cf. Fig. 1 of Liu et al., 2017) from two vegetative tillers of D. glomerata and 16 vegetative tillers of L. perenne, P. pratensis and P. pratense, one half of a leaf blade of T. officinale (with the latter severed along, but not including, the midvein) and two trifoliate leaves of T. repens. This protocol ensured collection of the entire within-leaf evaporative 18O gradient of all sampled leaf blade tissue of the different species. Stem (xylem) samples comprised the midvein of T. officinale, the petioles of the two T. repens leaves and the basal part of the vegetative grass tillers, except for the outermost part that was removed as it could have been subject to evaporative enrichment (cf. pseudo-stem in Fig. 1 of Liu et al., 2017).

Atmospheric moisture was collected by pumping ambient air through a glass coil immersed in a dry ice–ethanol mixture at a flow rate of 1 L min-1 over periods of 2–6 h around noon. Groundwater was sampled from a well located at about 100 m upstream of the ground water flow beneath paddock no. 8.

All plant and soil samples were immediately transferred to 12 mL Exetainer vials (Labco, High Wycombe, UK), sealed and covered with Parafilm. All samples were stored in a freezer at approx. -18 ∘C until water extraction. Water was extracted for 2 h using a cryogenic vacuum distillation apparatus with sample vials placed in a water bath with a temperature set to 80 ∘C (Liu et al., 2016).

Oxygen isotope composition was expressed in per mil (‰) deviation relative to a standard: 1δ18O=(Rsample/Rstandard-1), where Rsample and Rstandard are the 18O/16O ratios of the sample and the V-SMOW standard (Vienna Standard Mean Ocean Water). Samples collected between 2007 and 2012 were analysed by cavity ring-down spectroscopy using previously described procedures (Liu et al., 2016). Water samples collected in 2006 were analysed with an IsoPrime isotope ratio mass spectrometer interfaced with a multi-flow equilibration unit (both GVI, Manchester, UK). Each sample was measured against a laboratory standard gas, which was previously calibrated against secondary isotope standards (V-SMOW, V-SLAP and V-GISP). Heavy and light laboratory water standards, which spanned the range of δ18O values in the dataset, were analysed every five samples. Analytical uncertainty was 0.2 ‰. δ18O measurements obtained by cavity ring-down spectroscopy were linearly related to those obtained by isotope ratio mass spectrometry (n=176; R2=0.99). In a previous study, we found no difference between the results from spectroscopy-based and pyrolysis-based measurements performed using a TC/EA HTC coupled to an isotope ratio mass spectrometer (see Liu et al., 2017).

The isotope-enabled soil–plant–atmosphere model MuSICA (Ogée et al., 2003, 2009; Wingate et al., 2010; Gangi et al., 2015) was parameterised for the studied grassland based on data collected at the site or taken from the literature (for details and parameter values, see below and Methods S2 and Table S1).

The model was validated with latent energy flux (LE) data obtained from an eddy covariance station (EC) at the site. According to that comparison (Fig. S1), MuSICA estimates were unbiased (LEMuSICA=0.997LEEC; R2=0.59). Further, we compared MuSICA predictions of total plant-available soil water (PAW, mm) in the entire topsoil with PAW modelling and data for the same site presented in Schnyder et al. (2006). For the 2007–2012 data, this yielded the relationship PAWMuSICA=0.99PAWSchnyderetal.2006+7.8(R2=0.83).

Although the MuSICA model is capable of simulating δ2H of water pools in the soil–plant system, we excluded those data in the paper, as (1) we are primarily interested in the processes leading up to the δ18O of cellulose, (2) we had noticed discrepancies in the model–data agreement for 2H/1H indicating fractionation (including a surface effect on 2H/1H of soil water at the experimental site; Chen et al., 2016) that are currently not accounted for in the model, and (3) we did not want to overload the paper with extra figures and discussion. Issues of 2H/1H fractionation of water including data from this experimental site will be addressed in a separate paper.

A sensitivity analysis was conducted in order to quantify the responsiveness of predicted midday δ18O of leaf, stem and soil water to plant morpho-physiological parameters that were expected to affect those predictions based on theoretical considerations and/or observed parameter variation at the site. As the leaf water enrichment submodels are embedded in the process-based model MuSICA, the effect of parameters not included in the leaf water δ18O models per se could be evaluated. Based on the ceteris paribus principle, the sensitivity was tested by varying one parameter while keeping all other parameters the same as in the standard MuSICA parameter set (Table S1). For a sensitivity run, the parameter was not decoupled from the equations in MuSICA; hence changing one parameter value at the same time affected all equations containing this parameter and all dependent variables. Parameter effects (sensitivities) were quantified by two variables: (1) the mean sensitivity relative to the reference run, obtained as the mean differences from the reference run as (∑i=1n(δsens,i-δref,i))/n, with δsens,i the δ18O of a given water compartment (leaf, stem, or soil at 7 or 20 cm depth) in a sensitivity run and δref,i that in the reference run, for a day i; and (2) the standard deviation of the sensitivity, obtained from the differences between δsens,i and δref,i. The latter illustrated how strongly the effect of a parameter varied between sampling days and hence how strongly it depended on the conditions encountered on one specific day. Thus, the sensitivity variables reported if changes in parameter values caused systematic/general effects (shown by the mean sensitivity), or cancelling effects (shown by the standard deviations of the sensitivity), or combinations, or lack of the two.

The high and low parameter values for the sensitivity analyses were chosen according to the range observed for grasses or grassland species, as reported in the literature or observed at the site (see Supplement). Values for individual parameters of the sensitivity analysis were set at -0.20 and 0.50 for φ, 1 or 12 mol m-2 for leaf water content (W), 7 or 25 for the slope of the BWB model (mgs), 0 or 193 mmol m-2 s-1 for the intercept of the BWB model (g0), 0.6 or 3.8 m2 m-2 for leaf area index (LAI), 3.6 or 11.7 cm for canopy height (hcanopy), 20 or 140 µmol m-2 s-1 for the maximum rate of carboxylation at 25 ∘C (Vcmax), 32 or 224 µmol m-2 s-1 for potential rate of electron transport at 25 ∘C (Jmax), and 0.08 or 0.265 m for the mean of the vertical root distribution (μroot). Vcmax and Jmax were altered in tandem to keep the ratio Jmax/Vcmax at 1.6 (Medlyn et al., 2002), the same as in the standard simulation (Table S1). Apart from those plant morpho-physiological parameters, the effect of alternative submodels for the liquid and vapour effective diffusivity in the soil was tested by replacing the Moldrup formulation by the Penman one. In addition, we investigated the effect of using uncorrected IsoGSM-predicted δ18Orain and δ18Ovapour data instead of local isotopic data (gap-filled with offset-corrected IsoGSM data; see Sect. 2.4.1) for the isoforcing of MuSICA. This served to illustrate the usefulness of having local rainwater δ18O data.

For comparison of predicted and observed data, we calculated the mean bias error (MBE=P‾-O‾, where P‾ is the mean predicted value and O‾ the mean observed value) between observed and predicted δ18O (or Δ18O), the mean absolute error (MAE=((∑i=1n|Pi-Oi|)/n), where Pi is the predicted and Oi is the observed value at time i, and n is the number of values; Willmott and Matsuura, 2005), and R2 values.

Simple and multiple linear regression analyses and Student t tests were performed in R, version 3.4.2 (R Core Team, 2017), and RStudio, version 1.1.383 (RStudio Team, 2017).

Growing season rainfall amounts and distribution differed between years, with total precipitation in the main growing period (May to August) varying between 321 mm (2006) and 514 mm (2010) (Fig. 1a). The mean δ18Orain signal tended to increase in the first half of the vegetation period and decrease later in the season (Fig. 1b). However, individual rain events sometimes differed markedly from the mean pattern, with excursions of up to +4.5 ‰ and -6.2 ‰ relative to the mean of the same month (Fig. 1b). The δ18Ovapour signal followed similar mean trends (Fig. 1c) and exhibited a significant correlation (P<0.001) with the δ18O of the previous rain event.

The observed δ18Osoil was generally more enriched at 7 cm than at 20 cm belowground (Table 1; Fig. 2a, b). This relative enrichment with shallower depth was particularly large in the first half of the vegetation period and averaged 1.7 ‰ in the entire dataset. The total observed range of δ18Osoil differed somewhat between the two depths and was 7.8 ‰ at 7 cm, i.e. 16 % greater than at 20 cm (Table 1).

In most years, δ18Osoil followed the rain pattern and increased during the course of the vegetation period at both depths (Fig. 2a, b). This increase was generally more pronounced at 7 cm than at 20 cm. Overall, the seasonal patterns of δ18Osoil were quite dynamic, with considerable differences between individual years.

MuSICA simulations with the standard parameterisation (Table S1) predicted the multi-seasonal dynamics of δ18Osoil well (Fig. 2a, b) except in 2006 when local data of δ18Orain were not available for the isoforcing (Fig. 1b) and δ18Orain data were taken from the global atmospheric model IsoGSM, once corrected for the mean model–data offset (Figs. S2–S4). The seasonal trends and monthly fluctuations of observed δ18Osoil were reproduced with relatively small error (MAE of 1.1 ‰ and 0.8 ‰ at 7 and 20 cm, respectively). Also, the bias was small as MuSICA overestimated δ18Osoil by 0.8 ‰ and 0.5 ‰ at 7 and 20 cm, respectively.

Volumetric soil water content (SWC) predicted by MuSICA using the standard parameterisation (Table S1) exhibited strong seasonal and inter-annual variations. With SWC values (in m3 m-3) expected to vary between 0.19 (permanent wilting point) and 0.46 (field capacity), a SWC of less than 0.25 at 7 cm belowground corresponds to <25 % of the maximum plant-available water at this depth and is therefore a good indicator of edaphic drought. Each year, soil moisture at 7 cm fell below this threshold, but with a timing that differed from one year to the next (Fig. 1d).

Observed δ18Ostem generally matched and followed that of δ18Osoil at 7 cm, independently of SWC, season and year (Figs. 2b, c, 3a and S10). Conversely, the relationship between δ18Ostem and δ18Osoil at 20 cm was generally weak, exhibiting large scatter and a significant offset between δ18Ostem and δ18Osoil at 20 cm for most of the data (Fig. 3c). Remarkably, for 90 % of all days on which the soil was classified as “dry” (predicted SWC<0.25), δ18Ostem was still closer to δ18Osoil at 7 cm than to δ18Osoil at 20 cm.

Barnard et al. (2006) showed that the δ18O of (pseudo-)stem water in grasses is very close to that of the water taken up by the root systems of grasses (see also Liu et al., 2017), meaning that root water uptake operates without 18O isotope fractionation. MuSICA simulations were based on this assumption and reproduced very similar relationships between δ18Ostem and δ18Osoil to those observed at both depths, with similar R2, MBE and MAE (Figs. 2–3), thus showing a close agreement between observed and predicted data. Importantly, the close correspondence of δ18Ostem with δ18Osoil at 7 cm depth was not affected by changes in SWC predicted by MuSICA (Fig. 3). Again, the strongest disagreement between predicted and observed δ18Ostem occurred in 2006 (Fig. 2c), when observations of local δ18Orain were unavailable.

Midday leaf water δ18O (δ18Oleaf) exhibited by far the greatest observed δ18O variations in the entire dataset (Table 1). Also, δ18Oleaf was unique in the way that it did not exhibit a general trend during the vegetation period (P=0.5; right panel in Fig. 2d). As on average δ18Ostem increased over the vegetation period while δ18Oleaf did not, Δ18Oleaf exhibited a significant decreasing trend over the vegetation period, with a decrease of 0.5 ‰ per month (P=0.01; right panel in Fig. 2e), in parallel with the increasing trend of relative humidity over the growing season (data not shown). Conspicuous short-term, parallel increases/anomalies of δ18Oleaf and Δ18Oleaf (i.e. changes in δ18Oleaf largely independent of variations of δ18Ostem) occurred occasionally in different years, e.g. in spring of 2008, late spring and early fall of 2009, and early summer of 2010.

Predictions of Δ18Oleaf with MuSICA agreed best with observations using the two-pool model with φ=0.39 (R2=0.42; Table 2) in the standard MuSICA parameterisation. This result was robust for different soil water conditions. Unbiased predictions of Δ18Oleaf were best obtained by decreasing φ by 0.03 (i.e. setting φ to 0.36) under dry soil conditions (SWC<0.25) and increasing it by 0.01 (i.e. setting φ to 0.40) under moist soil conditions (SWC≥0.25), but this was an insignificant adjustment that did not change the overall coefficient of determination between observed and predicted Δ18Oleaf.

The agreement between observed and predicted Δ18Oleaf was always weaker when using the Péclet model. Fixing the effective path length (L) at a certain value led to predictions that were systematically biased for either dry or moist soil conditions (Table 3). Unbiased predictions of Δ18Oleaf in conditions of different SWC were only obtained when increasing L (from 0.162 to 0.235 m) for dry soil conditions and decreasing L for moist soil conditions (from 0.162 to 0.142 m).

MuSICA predictions of δ18Oleaf and Δ18Oleaf obtained with the standard parameterisation agreed well with observations at all timescales (Figs. 2d, e, S7 and S9), with low or no bias (MBE of 0.3 ‰ and 0.0 ‰, respectively) and a MAE for δ18Oleaf of 1.6 ‰, i.e. 10 % of the total variation of δ18Oleaf in the entire dataset (Tables 1, 2). Also, the relationship between modelled transpiration rate and the proportional difference between the observed Δ18Oleaf and Δ18O predicted by the Craig–Gordon model (Fig. S11) was non-significant, revealing no evidence of a Péclet effect. This was also true when investigating that relationship with a subset of the data that included only the leaves that exhibited near-steady-state 18O enrichment. This subset was estimated using model output to identify the times when near-steady-state conditions were most likely and included about half of the data (results not shown).

Multiple regression analysis demonstrated significant effects of air relative humidity (P<0.01) and SWC (P<0.05) on both observed and predicted Δ18Oleaf (Table 4). Δ18Oleaf increased with decreasing air relative humidity and SWC (Figs. 4a, b and 5a, b). The interaction effect of air relative humidity and SWC was close to significant for both observed (P=0.080) and predicted (P=0.073) Δ18Oleaf (Table 4). The effect of dry soil conditions on Δ18Oleaf was most evident at low air humidity (Figs. 4a, b and 5a, b) and was connected with a decrease in canopy conductance (gcanopy) (Fig. 5c).

The modelled dependence of transpiration on air VPD (vapour pressure deficit, the climatic driver of transpiration) was strongly modified by SWC (Fig. 4c). High air VPD drove high transpiration rates only under wet soil conditions (SWC≥0.25).

Increasing (decreasing) the proportion of unenriched leaf water (φ) and leaf water content (W) led to a strong reduction (increase) in δ18Oleaf (Fig. 6a, b). These changes in leaf-level parameters had no effect on δ18Osoil or δ18Ostem. Alterations of stomatal responsiveness (mgs), minimum conductance (g0), LAI or maximum carboxylation (Vcmax) and electron transport (Jmax) rates had similar directional effects (reflected by the mean sensitivity in relation to the standard simulation) on predicted δ18O of soil, stem and leaf water. However, the strength of the effects differed for the different ecosystem water pools (Fig. 6). Stronger effects were found on δ18Oleaf and δ18Osoil at 20 cm, compared to δ18Ostem or δ18Osoil at 7 cm that tended to vary in close harmony. Generally, a change in the parameter value caused an opposite change in the predicted δ18O of a given pool. Moreover, these parameters caused strong cancelling effects, evidenced by large standard deviations of the sensitivity, particularly for δ18Oleaf. The sensitivity of δ18Osoil to plant morpho-physiological parameters was related to the effect of those parameters on plant transpiration rate (not shown), which in turn altered the residence time of soil water at the lower depth. For example, lower Vcmax and Jmax values, not accompanied by a change in stomatal responsiveness mgs, implied a decrease in transpiration rate and consequent increase in the percolation of growing season rain water to the lower part of the soil profile (Figs. 7a and 8). In comparison, the 18O-depleted (winter) signal persisted longer in the lower profile at intermediate (Fig. 7b) or high (Fig. 7c) Vcmax and Jmax, as linked higher transpiration rates caused greater drying of the topsoil and reduced replenishment of deeper soil layers by summer rainfall.

Apart from LAI, other shoot characteristics, such as canopy height (Fig. 6f), leaf inclination, shoot shelter factor, leaf size and shoot size (not shown), had a very small or no effect on predicted δ18Oleaf, δ18Ostem and δ18Osoil.

The formulation of the water vapour diffusivity through the soil matrix (Fig. 6i) and the average rooting depth (Fig. 6h) affected δ18Osoil (and more strongly so at the lower depth), while the effect on δ18Ostem and δ18Oleaf was much weaker. Not accounting for the pore-size soil particle distribution parameter in the soil diffusivity formulation caused a greater overestimation of δ18Osoil, especially at 20 cm belowground where the MBE reached 1.3 ‰, compared to 0.5 ‰ in the standard run. Shifting the root distribution closer to the soil surface had little effect on δ18Osoil at both depths. Conversely, shifting it towards greater depth (Fig. S8) led to an overestimation of δ18Osoil, especially at 20 cm (Fig. 6h), and increased MAE in the relationship between δ18Ostem and δ18Osoil at both soil depths (not shown).

We also tested the effect of the choice of the water isotope forcing of MuSICA (δ18Orain and δ18Ovapour). In general, the agreement between predicted and observed ecosystem water pool δ18O was much better when MuSICA was forced using locally measured δ18Orain and δ18Ovapour data (Fig. 6j). The MBE for the δ18O of the different water pools was 3.1 to 6.7-fold greater when using the IsoGSM-based isotope forcing, and the MAE was 1.5 to 2.6-fold higher.

An isotope-enabled, process-based soil–plant–atmosphere model, MuSICA, generated realistic predictions of multi-seasonal dynamics of δ18O in soil, (pseudo-)stem and midday leaf water, as well as of the 18O enrichment of leaf water in a drought-prone temperate grassland ecosystem. Throughout the vegetation periods of seven consecutive years (1) model bias (MBE) was low, (2) the range of δ18O variations of the different ecosystem water pools was similar in the predictions and observations, and (3) prediction error (MAE) was less than 15 % of the total observed range of δ18O in the different ecosystem water pools and about twice the size of the MAE for the duplicate samples of the different pools. The relationships between observed Δ18Oleaf and variables related to the water cycle such as SWC, air relative humidity, transpiration and canopy conductance were well captured by the model. Although MuSICA is a detailed and locally parameterised model, this general agreement between model predictions and observations is remarkable given that model parameters describing the relevant physical features or functional relationships of soil and vegetation were held constant with one single value for the entire mixed-species ecosystem. This is a striking outcome given that predicted δ18O were found to be sensitive to several (but not all) plant morpho-physiological parameters (Fig. 6). The greater scatter in the observed relationship between Δ18Oleaf and relative humidity compared to predictions (Fig. 4) likely resulted partly from sampling effects and error. Sampling effects could include small-scale spatial variation of soil properties, or spatio-temporal variation of LAI, nutrient levels and root distribution, a regular feature of grazed grassland (e.g. Schnyder et al., 2006, 2010). Also, Webb and Longstaffe (2003) observed differences of several per mil in δ18Osoil in the top 5 cm over distances of about 10 m in a sand dune grassland. Such spatial variations would inherently cause greater scatter in the observations compared to the model predictions.

Prediction of δ18Ostem at a given point in time is a real challenge, as δ18Ostem is influenced by numerous factors, including the temporal distribution of rainfall amounts and its associated isotopic composition, transport and mixing of rainwater with soil water, the depth distribution of root water uptake in the soil, and soil evaporation. These ecohydrological processes are described explicitly in MuSICA, and agreement between observations and predictions of δ18Ostem and δ18Osoil at 7 and 20 cm depth indicates that MuSICA is capable of simulating these ecohydrological processes including 18O of the different water pools. The ability of the model to generate realistic predictions of the δ18O dynamics at different depths in the soil (within the zone of most active root water uptake and just below that zone) suggests strongly that the ensemble of parameters dictating the spatio-temporal dynamics of soil water contents (including emptying and refilling dynamics) was described well in the model. That interpretation was also supported by the sensitivity analysis. Importantly, a better agreement between predicted and observed δ18Osoil at 7 cm and δ18Ostem was obtained when the δ18O of meteoric water was taken from local measurements rather than given by the isotope-enabled atmospheric model IsoGSM (Fig. 6j). This result is not surprising given the significant spatial and temporal variation of rainfall at weekly and sub-kilometre scales (Fiener and Auerswald, 2009) and the comparatively large grid size of the IsoGSM model simulations (ca. 200 km ×200 km). Our model sensitivity analysis also revealed a better predictive power of the soil diffusivity formulation proposed by Moldrup et al. (2003) over that proposed by Penman (1940) to reproduce the observed isotopic composition of all the ecosystem water pools (Fig. 6i). This superiority was likely related to the effect of accounting for the soil-pore-size distribution parameter for describing the effective liquid water and water vapour diffusivity through the soil matrix and estimating this parameter from the soil water retention curve parameters measured at the site.

The comparison of observed δ18Ostem and δ18Osoil (Fig. 3a) strongly suggested that root water uptake occurred mainly at shallow depths (<20 cm) throughout the vegetation periods, largely independently of changes in SWC. This interpretation of observed data was based on comparison of δ18Ostem and δ18Osoil at two depths (7 and 20 cm) only, which provides limited spatial resolution and cannot inform precisely on the depth of root water uptake (Rothfuss and Javaux, 2017; Brinkmann et al., 2018). Such information can be improved by a locally parameterised, physically based, 18O-enabled ecohydrological model, as shown here. For instance, the standard MuSICA runs (Fig. 3b) indicated near-monotonous increases in δ18Osoil between 20 and 7 cm depth, matching well the observations in the majority of sampling dates (Fig. S13). Further, the simulations predicted a mean (uptake-weighted) depth of root water uptake above 15 cm in 90 % of all sampling dates, independently of SWC and observations of δ18Osoil (Figs. S12 and S13). Support came also from the MuSICA sensitivity analysis (Fig. 6h), showing that δ18Ostem was well predicted by the model only when root length density was maximum at shallow soil depth. The potential range of rooting depths is large in grassland, depending on site, species, climatic and management effects (Schenk and Jackson, 2002; Klapp, 1971). So, why would root water uptake be constrained to shallow depths in this drought-prone permanent grassland system? Several factors likely contributed: (1) the shallow topsoil overlying calcareous gravel (Schnyder et al., 2006); (2) the rapid shoot and root biomass turnover, which is associated with high phytomer dynamics leading to short leaf and root lifespan in intensively managed grassland (Schleip et al., 2013; Yang et al., 1998; Auerswald and Schnyder, 2009; Robin et al., 2010); (3) the high rates of shoot tissue (mainly leaves) losses that elicit a priority for assimilate (including reserve) allocation to shoot regeneration at the expense of the root system (e.g. Bazot et al., 2005); and (4) predominant placement of the root system near the soil surface dictated by the high need for nutrient interception and uptake (e.g. from excreta deposits), to compensate for the high rates of nutrient losses due to grazing (Lemaire et al., 2000). Importantly, (5) in a relatively high number of cases, the model predicted situations in which rainfall recharged mainly the topsoil, while SWC at depths >20 cm remained low (e.g. June–end of year 2006, April–October 2007, or May–end of year 2008; Fig. S12; see also below). Principally, however, factors (2)–(4) alone can explain why shallow rooting depth is a typical feature of intensively grazed grasslands (Troughton, 1957; Klapp, 1971). Also, Prechsl et al. (2015) did not find an increasingly deeper root water uptake upon soil drying in an alpine and a lowland grassland system in Switzerland. Similarly, grasses continued to rely on water in the uppermost soil layer during soil water scarcity in a mesic Savanna in South Africa, in which C4 grasses were growing together with saplings and trees (Kulmatiski and Beard, 2013), and in a tallgrass prairie in the US dominated by C4 grasses and C3 shrubs and forbs (Nippert and Knapp, 2007a, b).

Predictions of δ18Osoil, particularly below the main zone of most water uptake, at 20 cm, were influenced markedly by estimates of LAI and by changes in Vcmax, Jmax, and stomatal conductance responsiveness (mgs) or minimal value (g0). This resulted from the effect of those parameters on total canopy transpiration, which in turn altered the dynamics of soil water and hence of the mixing of 18O-depleted winter and 18O-enriched summer precipitation with soil water at different depths. For instance, an increase in transpiration rate caused by a high mgs led to a decrease in δ18Osoil at 20 cm during the course of the growing season and a growing divergence between observations and predictions, particularly in years with low growing season precipitation (data not shown). This was likely caused by the fact that 18O-enriched summer rain mainly recharged the upper soil layer in this scenario, as this had been desiccated extensively because of the higher transpiration resulting from the higher mgs. So, summer rains would contribute less to wetting of the lower profile. Conversely, if mgs was set to a low value, predicted δ18Osoil at 20 cm increased throughout the vegetation period. According to the same mechanism, the effect of mgs on δ18Osoil was negligible when growing season rainfall was high in 2010. The effects of changing Vcmax and Jmax, LAI and minimum conductance on predicted δ18Osoil at 20 cm were very similar to mgs, suggesting that these parameters acted via the same mechanism, that is canopy conductance for water vapour that is controlled largely by the (integrated) stomatal conductance of all leaves within the canopy. Thus, the effect of Vcmax and Jmax was likely indirect, resulting from altered assimilation rates impacting stomatal conductance.

The Δ18Oleaf data were well predicted with a two-pool model and a constant fraction of unenriched water in bulk leaf water (φ≈0.39). This model was valid for a wide range of atmospheric and soil water conditions in seven consecutive growing seasons. Inclusion of a Péclet effect reduced the closeness of fit between measured and modelled Δ18Oleaf under all environmental conditions. We did not know if putative between-species differences in leaf water dynamics and associated 18O enrichment, or any other morpho-physiological effects for example associated with leaf ageing, could have led to a missing correlation between the proportional difference between measured leaf water 18O enrichment and that predicted by the Craig–Gordon model (1-Δ18Oleaf/Δ18Oe) and transpiration rate. For these reasons, we explored this question with separate studies of L. perenne and D. glomerata, two species that also formed part of the present grazed grassland ecosystem. Again, these studies found no evidence for a Péclet effect and supported the two-pool model, as there was no relationship between the proportional difference between measured leaf water enrichment and that predicted by the Craig–Gordon model (1-Δ18Oleaf/Δ18Oe,ss) and transpiration rate in either L. perenne plants grown in a controlled environment at different relative humidities and water availabilities, or D. glomerata leaves measured using an online transpiration isotope method (Notes S2 and Figs. S14, S15). A two-pool model was also suggested by the diurnal time courses of δ18Oleaf in this grassland (Fig. S7) and in a broadleaf and a coniferous tree species (Bögelein et al., 2017).

When interpreted with the Péclet model, the two-pool model implies a constant Péclet number and inverse variation of transpiration rate and effective path length (L). Dynamic changes in L in response to varying transpiration have been noted before, mainly in controlled conditions, and interpreted in terms of changing contributions of different paths (symplastic, apoplastic and transcellular) of water movements to the stomatal pore (Barbour and Farquhar, 2003; Kahmen et al., 2008; Song et al., 2013; Loucos et al., 2015; Cernusak et al., 2016). Increases in L in response to drought, as suggested in this work, have also been observed previously in Vitis vinifera by Ferrio et al. (2012) and were connected with variations in leaf lamina hydraulic conductance.

In principle, failure to detect a Péclet effect could be related to the presence of major veins and associated ground tissue of the grass leaves (Holloway-Phillips et al., 2016) or errors associated with non-steady-state effects on 18O enrichment of bulk leaf water (Cernusak et al., 2016). However, MuSICA predictions of Δ18Oleaf did account for non-steady-state effects and were generally consistent with observed Δ18Oleaf. The φ value used in our simulations is in the upper range of φ values reported for grasses. Liu et al. (2017) observed species-specific φ values ranging from -0.05 to 0.43 in two C3 and three C4 grasses, with no obvious effect of vapour pressure deficit on φ. Gan et al. (2003) presented φ values between ca. 0.16 and 0.41 in maize, with lower values coming from leaves with the midvein removed. Considering a similar effect of vein removal would move our observed φ to about 0.2. Such a value of φ for grasses is very similar to the mean φ reported for a wide range of non-grass species by Cernusak et al. (2016).

The strong response of Δ18Oleaf to air relative humidity has been observed and discussed previously (e.g. Farquhar et al., 2007; Cernusak et al., 2016), in addition to soil moisture (Ferrio et al., 2012). We are not aware of a previous study that disentangled the separate effects of atmospheric and soil humidity on Δ18Oleaf, either in field or controlled conditions. Notably, the responses observed in our work were corroborated by theoretical predictions as implemented in MuSICA. Modelled transpiration rate and stomatal conductance were greatly reduced under dry soil conditions, leading to higher kinetic fractionation αk (Eq. 3) but lower α+ (Majoube, 1971) and relative humidity h, because of the warmer leaf temperatures. The net effect was a greater Δ18Oleaf predicted by MuSICA under dry soil conditions, in agreement with observations. This demonstrated that other vegetation parameters that affected the 18O enrichment in our sensitivity analysis (e.g. the unenriched fraction φ or the effective mixing length L, leaf water content W or LAI), but were not considered drought-sensitive, did not seem to be the main drivers of the enhancement of Δ18Oleaf during edaphic drought.