2.1 Experiment description: The Nutrient Network (NutNet)

We used total and functional group aboveground net primary production (ANPP) data from 10 European grassland sites of the Nutrient Network, a standardized nutrient addition experiment (Borer et al., 2014; 2017), sampled from at earliest 2009, until 2018 (Figure 1—n = 3–5 plots per treatment and 3–9 years per site; Table 1). Grasslands were spread over Europe: four sites in Germany, two in Finland, and one each in Ireland, France, Portugal, and Switzerland. These sites were selected because all had actively running experiments during the European drought of 2018 and had more than 1 year of data.

The sites were located between 38.83 and 69.06°N, with elevations ranging from 40 to 2320 m a.s.l., mean annual temperatures from −3.3°C to 16.6°C, and mean annual precipitation from 521 to 1320 mm. Grassland types varied, with temperate grasslands being the most common. Five of the sites were managed with varying mowing or grazing regimes. Site-specific information, such as grassland type, climate, sampling years and dates, soil organic matter content, and soil texture are given in Table 1.

The standardized treatment protocol involves a full-factorial annual N (10 g m⁻² year⁻¹), P (10 g m⁻² year⁻¹) and K (10 g m⁻² year⁻¹) addition experiment on 5 m × 5 m grassland plots. Along with K, a mixture of 100 g/m² of micronutrients (1.5% Mg, 14% S, 0.2% B, 1% Cu, 15% Fe, 2.5% Mn, 0.05% Mo, 1% Zn) was added but only in the first year of treatment. At some of the sites, nutrient treatments were combined with a treatment of grazing exclusion (fencing), and/or “open roofs” that were used as controls of additional drought experiments. Data from such treatments were included in the present study to assess the robustness of our results across, for example, varying levels of grazing pressure.

Total biomass was sampled and separated into current year's graminoids (i.e., grasses, rushes, and sedges), leguminous and nonleguminous forbs, bryophytes, lichens, and woody vegetation once per year at most sites, usually at peak season, to estimate ANPP (g m⁻² year⁻¹). At sites bayr.de and jena.de, biomass was sampled both at the beginning of summer and at the end of the growing season; biomass values were then summed. Together with climate, background nutrient status, N deposition, species composition, etc., the sampling regime thus represents another source of uncertainty. Given that clear patterns emerged in spite of these differences among sites points toward a strong robustness of our findings. Two randomly placed 0.1 m² quadrats were harvested on each plot, sorted to functional groups and dried at 60°C for at least 48 h.

2.2 General approach

To make responses and explanatory variables comparable across the sites, we normalized the annual biomass production and components per plot and year (nANPP) by taking the natural logarithm of each year's biomass divided by the mean value of all sampling years in that plot (n = 3–9 years per site; Extended Data Table 1). nANPP is similarly interpretable as the log-response ratio, a common metric used in meta-analyses, with values >0 indicating higher-than-average, and <0 lower-than-average biomass production. For the explanatory variable, we estimated a soil drought index (I_s) that reflects the magnitude of drought experienced over the entire growing season at a site, based on daily weather (precipitation, potential evapotranspiration [PET]) and soil (texture, organic matter) data (see Section 2). For calculation of I_s, we used the threshold proposed in Granier et al. (2007) and Vicca et al. (2012; relative extractable water [REW] = 0.4). In the absence of drought (i.e., REW > 0.4), I_s equals zero. The longer and more intense a drought, the higher I_s. For inter-site comparison, we expressed this as distance to the average by subtracting I_s from the inter-annual mean (Diff I_s = I_s − mean I_s for that site taken over all biomass sampling years; Diff I_s > 0 means more drought stress than average, Diff I_s < 0 means less drought stress than average). While actual soil moisture data were not available for most sites, such index based on “potential” soil moisture under unconstrained evapotranspiration offers the advantage that it takes into account among-site variation in the capacity of soils to store and supply water to biota. This soil influence is important for inter-site comparison but is ignored in analyses focused solely on water inputs (Vicca et al., 2014). We complemented our analyses using the soil drought index with results based on data of precipitation and the water balance (precipitation – PET) presented in Figure S2 and Table S1, which yielded qualitatively similar results.

2.3 Weather data and soil drought index

Daily minimum and maximum temperatures (T_min and T_max), and precipitation were measured on-site, or were collected from nearby weather stations which had data archived at the Global Historical Climatology Network (NOAA National Centers for Environmental Information). The growing season was then defined per calendar year as the period between the first span of at least 6 days with average temperature >5°C, and the first span of at least 6 days with average temperature <5°C (Zhang et al., 2018). Growing season degree days (GDD; °C days), precipitation (Pr; mm), water balance (WB; mm) and the soil drought index (I_s) were then calculated from the start of the growing season until the day of sampling (the last sampling day was used where two biomass samplings per year occurred), to better represent the weather as experienced by the biota compared to annual full growing season meteorological data.

Growing season degree days was calculated by summing average daily temperature data minus five (average taken as (T_min + T_max)/2, days <5°C count as “0”), from the start of the growing season until the day of sampling (Mcmaster & Wilhem, 1997). Pr was calculated by summing daily precipitation (Pr_day) over the same period. WB was defined as the difference between Pr and PET, with PET_day estimated from the Hargreaves equation, using latitude, day of the year, T_min, and T_max (Yates & Strzepek, 1994). Calculations of the daily soil drought index (I_s,_day) required WB_day, along with soil texture and soil organic matter content, which the estimates of soil permanent wilting point (PWP %) and field capacity (FC %) were based on, following Saxton and Rawls (2006). I_s was then calculated by assigning a stress threshold for soil moisture (at REW of 0.4), and integrating daily values below this threshold over the growing season until the day of sampling (Granier et al., 2007; Vicca et al., 2012):

(assuming a main rooting zone of 20 cm depth, e.g., Kübert et al., 2019);

where TEW is the total extractable water;

(TEW_day-1 for January 1 was taken as TEW_max each year,

e.g., Coleman & Jenkinson, 1996; TEW_day ≥ 0 and TEW_day ≤ TEW_max);

where SWC is the volumetric soil water content;

The “potential” SWC data (influenced by PET and Pr) were then compared to actual SWC measurements (influenced by actual evapotranspiration, Pr, horizontal and vertical water transport, etc.) from 2018 in Control plots at site free.de (Table 1), the only site and treatment with continuous SWC measurements over a full season. Our SWC estimates correlated strongly with actual SWC (Pearson's r = 0.81 for 5 cm soil depth; r = 0.73 for 20 cm soil depth; Figure S1).

In line with previous studies (Vicca et al., 2012), we suggest that drought stress indices from a soil moisture perspective (I_s) better grasp the actual drought as experienced by the biota than solely meteorological data (Pr and WB). As advantages, we note for example (i) increases in the stress index only during periods when soil moisture is lower than a predefined threshold below which plants likely experience drought stress, (ii) higher suitability for inter-site comparison by taking into account differences in soil water holding capacity, (iii) avoiding a too strong influence of extreme precipitation events (e.g., a short heavy rainfall event cannot fully replenish soil water because some water is for example lost via runoff), and (iv) reduced sensitivity to slightly varying sampling times across years within one site (e.g., later than other years end-of-season sampling could be associated with higher precipitation compared to other years simply because of the longer growing season, while integrated soil moisture drought stress over the growing season and collected biomass may have been unaltered).

2.4 Data analyses

3.1 Graminoid response dominates nitrogen addition effect on community biomass production

Nutrient additions can alter functional group composition and thereby influence the drought sensitivity of the entire plant community. Nutrient-induced shifts in functional group dominance were indeed observed: fertilizer addition increased aboveground grassland biomass production and this response was mainly driven by the response of the graminoids (Figure 2; ANPP_community: F_7,162 = 7.34; p < 0.001***); graminoids exhibited a clear, positive response to treatments that included N (ANPP_graminoid: F_7,282 = 20.71; p < 0.001***), with the effect increasing when P and/or K were additionally included in the fertilizer mix. Forbs showed no nutrient addition effect on biomass production (ANPP_forb: F_7,123 = 1.00; p = 0.43), such that the dominance of graminoids in the herbaceous community increased under NK, NP, and NPK addition (ANPP_{graminoid/herb}:  = 72.41; p < 0.001***). It has been suggested that strong positive effects of nutrient addition on ANPP_graminoid, but not ANPP_forb can be ascribed to promoted growth of dominant grass species with resource-acquisitive traits (Blanke et al., 2012; La Pierre et al., 2016). Such grasses can then suppress a positive response of the forb functional group through increased competition for light (Hautier et al., 2009).

A shift occurred in the contribution of legumes versus nonlegumes to overall ANPP_forb: ANPP_legume decreased under nutrient additions that included N (F_7,252 = 7.14; p < 0.001***), while ANPP_nonlegume showed no significant overall tendency (F_7,249 = 1.62; p = 0.13). The advantage of legumes fixing atmospheric N₂ thus appears less relevant when N is ample (Fu & Shen, 2016), increasing the impact of competition for light on the legumes (Hautier et al., 2009).

Fencing (grazing exclusion) treatments were combined with Control and NPK treatments at eight sites and were included in this study to verify the generality of our results on the interaction between drought stress and nutrient addition across European grasslands varying in grazing pressure, climate, plant species composition, etc. Grazing exclusion did not affect the ANPP of individual functional groups (p > 0.10) but had an overall positive effect on ANPP_community (F_2,384 = 10.64; p < 0.001***), independent of NPK addition (F_2,377 = 1.31; p = 0.27; Table 1). This result is in line with many earlier synthesis studies that found no interaction effect between nutrient addition and herbivory on community biomass (Borer et al., 2017; Gruner et al., 2008), despite that the interaction has been observed at the level of species composition and biodiversity (Borer et al., 2017; Brandt et al., 2019).

3.2 Drought effect on grassland ANPP depends on nutrient availability

Inter-annual analyses revealed that the soil drought index exhibited a significantly stronger negative effect on nANPP_community for NP and NPK addition than for other nutrient treatments (Figure 3a). In other words, NP and NPK addition aggravated drought effects on grassland ANPP. Also the water balance (precipitation minus PET) had a significant positive influence for NPK addition (+0.08 ± 0.04/100 mm), but not for other treatments (Table S1). Our results thus illustrate how NPK (and NP) addition can exacerbate the drought effect on aboveground biomass production.

For the year 2018, when summer was extremely hot and dry at most sites, a strong negative relationship (−0.13 ± 0.05/10 drought index units) was found between the soil drought index and nANPP_community, irrespective of nutrient treatments (Figure 3b; Table S2). The overall drought effect was even stronger than in the NPK treatment of the inter-annual analysis. This observation may be ascribed to the particular extremity of the summer of 2018 (Buras et al., 2020), which seemed not completely grasped by the soil drought index (cf. the horizontal position of some points in Figure 3b). Substantial suppression of biomass production occurred at the dry sites, average production at a few sites with moderate drought and above-average biomass production at site comp.pt, which exhibited an anomalously wet growing season that year (Figure 1; Table 1). Drought effects did not depend on the nutrient treatment here, but this may be due to the limited sample size of this partial dataset.

Our continental-wide multisite comparison based on inter-annual data suggests a negative interaction effect between drought and nutrient availability. This interaction effect was clear not only on absolute, but also on relative-to-average values of aboveground grassland production (i.e., under drought, nANPP not only decreased more in NPK than in Control in absolute terms, but also in relative terms, suggesting a strong interactive effect). The negative interaction effect between drought and nutrients was most evident when comparing treatments including the most limiting nutrient (that is, N; Figure 2; Table S1) with ambient nutrient conditions. Our findings thus support previous, usually smaller scale or single-site studies suggesting that fertilization aggravates the impact of drought.

The negative interaction effect between drought and nutrients was reported in earlier studies focusing both on multi-species communities (Bharath et al., 2020; Kübert et al., 2019) as well as single species/monocultures (studies in grasslands but also other ecosystems; Friedrich et al., 2012; Meyer-Grünefeldt et al., 2015). A negative drought × nutrient interaction at the individual and population (i.e., within-species) level can, for example, occur through reduced root–shoot ratios under nutrient addition (Blanke et al., 2012) and increased pre-drought evapotranspiration associated with faster soil moisture depletion (Kübert et al., 2019). Such effects may counterbalance potentially positive nutrient influences on ANPP such as higher production of osmoprotectants (Saneoka et al., 2004), or deeper absolute rooting depth (Gessler et al., 2017). Our study used data from grassland communities where such individual species responses to drought, nutrient availability, and their interaction act concurrently with interspecific processes (e.g., release from competition).

Greater drought impacts on nutrient-rich than nutrient-poor grasslands can result in a shifting balance among ecosystem services at the landscape scale. Fertilized and non-fertilized grasslands differ in the services they provide; fertilization is typically more intensive in systems with provisioning services such as (fodder) production as primary use, whereas no input grasslands provide supporting or regulating services such as biodiversity and pollination (Zhao et al., 2020). In a future with more frequent and intense droughts, uncertainty of provisioning grassland services may therefore increase more than that of supporting and regulating ecosystem services. Depending on society's demands and preferences (e.g., maintaining high average but highly variable yield vs. more stable regulating services), shifts in the balance between high versus low intensity land use can be made as a climate adaptation strategy through adjustments of agricultural policies.

3.3 Nutrient addition intensifies negative drought effects on graminoids, but not forbs

Differential responses of plant functional groups played a significant role in shaping the drought impact on community productivity. Contrary to our expectations, drought had no effect on forb biomass but a negative effect on graminoids under some fertilization regimes (Figure 4a vs. 4b; Table S1). The negative impact of drought on graminoids (nANPP_graminoid) was found in NK (−0.09 ± 0.06/10 index units or +0.18 ± 0.09/100 mm), NP (−0.13 ± 0.06/10 index units or +0.18 ± 0.09/100 mm), and NPK (−0.13 ± 0.05/10 index units or +0.18 ± 0.09/100 mm) treatments but not in Control, N, P, and PK treatments (Figure 4a; Figure S2c; Table S1). Increasing drought resulted in a reduction of the proportion of graminoid to herbaceous (graminoid + forb) biomass production (nANPP_{graminoid/herb}) in fertilized N, NK, NP, and NPK treatments (Figure 4c; Table S1). Using the data subset from the extreme year 2018 only, increasing intensity of drought across sites was also associated with a reduction in the production of graminoids, albeit only marginally significantly (−0.15 ± 0.08/10 index units), and independently of nutrient addition (Figure 4d; Table S2). Forbs again showed no significant response to drought.

A higher impact of drought on graminoids than on forbs, and therefore reduced proportions of graminoids under drought (Figure 4c), contrasts with earlier studies that suggested stronger drought effects on forb biomass production than on graminoid biomass production (Breshears et al., 2016; Hoover et al., 2014; Reynaert et al., 2020; Stampfli et al., 2018; but see Kübert et al., 2019). We suggest that different mechanisms are at play that can either predispose or protect different functional groups against drought impacts. While higher phenotypic plasticity under drought in some traits such as specific leaf area may put graminoids at an advantage (Wellstein et al., 2017), their resource-acquisitive traits such as a shallow rooting system predispose common graminoids to reduced drought resistance, as opposed to forbs with deeper (tap)roots (Berendse, 1982; Zeiter et al., 2016). Moreover, also high initial evapotranspiration may promote soil moisture depletion in the shallow rooting zone of the graminoids, further amplifying drought stress (Kübert et al., 2019). Consequently, a graminoid-dominated plant community can become even more drought-sensitive under nutrient addition and this way drought can thus counteract graminoid dominance in nutrient-rich grasslands.

Besides variation in the role of mechanisms regulating drought sensitivity, differences in time scale could also partly explain apparent discrepancies in graminoid versus forb dominance between our study and others. The immediate drought response (resistance) may differ strongly from the response in the longer term that also includes recovery (Ingrisch & Bahn, 2018), and resistance and recovery often show opposite patterns (Bharath et al., 2020; Karlowsky et al., 2018; but see Wang et al., 2018). With our approach, we linked grassland responses to meteorological data of the same growing season. As a consequence, these responses reflect drought resistance, plus some within-year recovery depending on when the drought occurred relative to the biomass harvest. Recovery, however, is a phenomenon that occurs not only within a growing season, but continues over a multi-year time scale (e.g., Bharath et al., 2020). The yet limited time scale of the European NutNet data series does not allow disentangling resistance and recovery on a multi-year scale. Possibly, graminoid dominance recovers over longer periods (Stampfli et al., 2018), although a long-term drought-induced shift to a community with higher forb abundance is also possible, as was exemplified by a mesic meadow in Switzerland (Stampfli & Zeiter, 2004). In fact, such persistent changes may become more likely when droughts increase in frequency and intensity as a consequence of climate change (Frank et al., 2015).

Drought reduced aboveground biomass production of graminoids under NP and NPK addition, but not under ambient nutrient conditions (Figure 3a; Table S1). The differential drought sensitivity of graminoids versus forbs exclusively in fertilized plots therefore suggests that, in addition to the plant trait effects discussed above, a nutrient-induced increase in evapotranspiration associated with the graminoid biomass production (Figure 2) may have increased graminoid susceptibility to drought. Such fertilization effects cannot be captured by our drought indices and actual soil moisture data were not available for most sites. Further research is therefore needed to verify if soil moisture is indeed depleted more quickly in fertilized plots (e.g., Harpole et al., 2007), increasing drought severity and eventually, exacerbating reductions in biomass production.

The observation that no drought effect was found on forbs in fertilized plots, as opposed to negative drought effects on graminoids (Figure 3), may also be explained by release of forbs from competition (Connell, 1961). Under addition of N (plus P and/or K), graminoid biomass production is typically increased but not that of forbs because of shading by the graminoids (Hautier et al., 2009; Figure 2). However, when drought occurs, water limitation can suppress the growth of graminoids, and potentially also that of forbs. Reduced graminoid growth can then reduce competition for light which balances out the negative drought effect on forbs. Hence, whether experiments find apparent higher drought sensitivity of graminoids compared to forbs in a community may depend on the strength of competition between the functional groups, which in turn depends on nutrient availability. Such links among nutrient addition, drought, and shifts in interspecific interactions should be investigated further in experiments that manipulate all three factors.

3.4 Legumes are more responsive to drought than nonleguminous forbs

While the analysis on inter-annual variation showed no effect of the soil drought index on nANPP_legume or nANPP_nonlegume (Figure 5a,b), the extreme year 2018 did reveal a significant positive effect of drought on legumes under NK and NPK addition, but not under other nutrient addition treatments (Figure 5c; Table S2). We suggest that the unexpected positive association between drought and legume production we observed may be an apparent effect, where drought suppressed biomass production of competitors (Figures 3 and 4), and in turn released legumes from competition (Connell, 1961). Interestingly, this effect seemed to be greatest where legume biomass was most suppressed (Figure 2).

Although not the primary topic of this study, we note that a similar, positive interaction effect was found between fencing and drought on legumes, in both the 2018 dataset (+0.4 ± 0.1/10 index units; Table S2) and the inter-annual analysis (+0.13 ± 0.04/10 index units; Table S1). While fencing can have positive effects on legume abundance if herbivores strongly prefer grazing on N-rich legumes (Ritchie & Tilman, 1995), such positive fencing effect was not found in our study. Instead, we found a tendency for a negative influence of fencing (Figure 2). Possibly, fenced plots that prohibit grazing can strengthen competition for light from fast-growing graminoids and thereby reduce legume productivity. Drought may have suppressed such light competition effect on graminoids, causing an apparent positive effect on legumes. Interactions of drought with fencing on other functional groups and on the whole plant community were absent or more ambiguous than interactions with nutrient addition (Tables S1 and S2).

Legumes did not show an overall response to inter-annual variation in the soil drought index (Figure 5a,b), but the method using the water balance suggested a positive water availability effect on this functional group (+0.3 ± 0.2/100 mm), and the method using precipitation indicated positive precipitation effects on legumes only under P (+0.6 ± 0.2/100 mm) and PK addition (+0.4 ± 0.2 /100 mm). Nonlegumes were unresponsive (Table S1). This result would be in line with our hypothesis of increased drought impacts (or at least positive precipitation effects) on legumes compared to nonlegumes, especially under nutrient additions without N. These are treatments where legumes are expected to profit most from their N₂ fixing capacity. Possibly, drier conditions suppressed symbiotic N₂-fixation, either directly because of water limitation (Serraj et al., 1999), or indirectly due to hampered P uptake (Mariotte et al., 2020) by legumes with typically high P demand (Mitran et al., 2018). The legume advantage under P and PK addition over other forbs (Figure 2) may thus reduce with decreasing water availability. Further research is needed to verify these speculations.

3.5 Beyond functional groups: Biodiversity, nutrient addition, and drought impacts

We found that grassland responses to soil drought stress depend on both nutrient availability and functional group composition. The importance of plant community composition in shaping the drought response is in line with the diversity–stability hypothesis. Much of the literature on this hypothesis refers back to the work of Tilman and Downing (1994), who found greater drought resilience in more species-rich plots of a long-term N addition experiment with gradient design. Species richness declined under N addition in that experiment, but the positive richness–resilience link was found to be stronger than the negative N addition–resilience or grass biomass–resilience links, emphasizing the importance of plant diversity for drought resilience. Later studies pointed to components beyond richness underlying the diversity–stability link, such as species asynchrony and trait diversity that both have a strengthening effect on the relationship (Craven et al., 2018; Weisser et al., 2017). Nutrient addition has been shown to reduce species asynchrony and therewith community stability, even in the absence of substantial changes in overall richness (Hautier et al., 2014, 2020). For future biodiversity-oriented research, we suggest that in-depth evaluations of responses of grassland productivity, species traits and asynchrony to the global changes drought, nutrient addition, and their interaction can shed more light on how nutrients and biodiversity shape community drought responses.