We evaluate stalagmites from two cave systems (Fig. 1) in northern Spain. La Vallina Cave is hosted in Carboniferous limestone of the Barcaliente formation and Cueva Rosa is hosted in the Escalada Formation. Modern vegetation above the cave sites is variably affected by anthropogenic disturbance, such as planting Eucalyptus crops, and maintaining pasture (including occasional burning for pasture regeneration). The climax vegetation of this area is temperature deciduous forest (Roces Díaz et al., 2015). La Vallina cave (43°24^′36^′′ N, 4°48^′24^′′ W; 70 m a.s.l. (above sea level) is on a northward-facing hillslope at a distance of 2.5 km from the coast, with rock thickness above the cave ranging from 5–30 m. Modern vegetation includes native Quercus ilex and Quercus rober as well as recently grown eucalyptus and pasture (see Kost et al., 2023, for overview) Cueva Rosa (43°26^′37^′′ N, 5°08^′25^′′ W; 121 m a.s.l.) features 50 m of bedrock above the cave, and a lower gallery is transited by a perennial cave stream which periodically floods some of the upper galleries (González-Lemos et al., 2015).

The annual precipitation at the cave sites is 1250 mm, based on the averages from 1970 to 2009 (data on annual rainfall from http://idebos.bio.uniovi.es/GeoPortal/Atlas/Pan19702009.html, last access: 28 April 2013). A water deficit emerges in summer months due to lower precipitation and increased evapotranspiration (for a review, see Kost et al., 2023). Mean cave temperatures are 12 °C (Kost et al. 2023), close to the annual average air temperature, with monthly air temperatures at the coastal site ranging from 9 to 20 °C (Fig. 1c).

Drip water monitoring in La Vallina cave found modern P concentrations to be below the analytical detection limit of 0.32 µmolL-1 (Kost et al.2023). The median bedrock P/Ca (determined as described in Kost et al. (2023) at La Vallina is 0.19 mmol mol⁻¹ and at Calabrez it is 0.08 mmol mol⁻¹ (Fig. 2). Because these estimates derive from opportunistically collected bedrock samples, we do not have a way to estimate the weighted mean bedrock composition at either cave setting.

Stalagmites GAE, GLO, and ROW span the 92 to 90 kyr interval and their carbon isotope record, corrected for in-cave degassing and prior calcite precipitation processes, record the cooling of the Greenland Stadial 22 (GS22) event as a positive shift in the δ13C_init (Stoll et al., 2023) (Table 1). Stalagmites GAR, NEI, GAE, GLD, and BEL span the transition from the Penultimate Glacial Maximum (PGM) to the Last interglacial (LIG); isotope and Mg/Ca records from GAR, NEI and GAE and GLO have been described (Stoll et al., 2022, 2015) and the δ13C_init record from GAR discussed (Kaushal et al., 2025) and the constraints for calculation of δ13C_init for GAR, GAE, GLD and BEL presented elsewhere (Stoll et al., 2023). In several of these stalagmites, previously published isotope and Mg/Ca records span longer time intervals than the P/Ca determinations presented here, because some previously published Mg/Ca were measured by ICP-OES for which we did not obtain reliable P measurements. As detailed elsewhere, trends in δ13C_init (and the index fCa, the fraction of calcium remaining in solution after prior calcite precipitation) may be calculated more reliably than the absolute values (Stoll et al., 2023), and here we focus on the interpretation of trends in these parameters. Sample chronology from U-Th dates derives from published studies (Stoll et al., 2013, 2022, 2023) (Table 1). Age models are plotted in Fig. S1. Median growth rates of the studied samples range from 4 to 36 µmyr-1. Age models for GAR include sectors with layer counted growth rates, and in counted sectors layer thickness suggests growth rates varying by 5 to 10-fold on the decadal to centennial scale. In this sample, layer thickness decreases during cold stadial events, and annual layers become too thin for counting between 134.1 and 132.3 ka, as well as during the LIG (Stoll et al., 2022). Age models for other stalagmites are based on the interpolation schemes of BCHRON (Haslett and Parnell, 2008) and STALAGE (Scholz and Hoffmann, 2011) or linear interpolation when the density of dates is low. With the exception of GAR, the available chronological resolution does not permit us to confidently identify variations in growth rates between U/Th dates, and in most samples growth rate variations cannot be identified at timescales shorter than several thousand years (typical interval between dates Fig. S1).

Stalagmite samples were drilled using a Sherline drill which digitally monitors sampling position. Five of the presented stalagmite datasets are based on drilling resolution that continually sampled all of the deposited stalagmite along the growth axis, using drilling resolution of 0.25 to 1 mm per sample (Table 1). These include most of the presented records covering the PGM-LIG and one of the four records covering the GS22. For these samples, each drilled increment typically integrates 30 to 230 years (Table 1). The other presented records sampled 1 mm of powder but at sample increments ranging from 2 to 5 mm, meaning that only 20 % to 50 % of the growth conditions were sampled. The typical interval between samples reflects 150 to 400 years for these datasets (Table 1). During GS22, each drilled sample from GAE, NEI and ROW integrates 30 to 70 years based on the median growth rates. For a 1.5 cm increment in GAE with higher sample resolution of 0.5 mm, this portion of the record has also been downsampled to reflect a constant sample spatial resolution. The impact of variable stalagmite growth rate on the variance in P/Ca is assessed in stalagmite GAR, which features a layer-counted age model from 136 to 128 ka (Fig. S2).

An aliquot of 330 to 400 µg of powders was dissolved in 350 µL of 2 % HNO₃. P/Ca and Al/Ca were analyzed with an Agilent 8800 QQQ-ICP-MS at ETH Zürich using the intensity ratio calibration (de Villiers et al., 2002) and in-house multielement standards of varying trace element to Ca ratios but with Ca concentrations matched to samples (400 ppm). We analyze ³¹P mass shifted in O₂ reaction mode (+16 mass units) to reduce ¹⁵N¹⁶O polyatomic interferences. We report data as the stalagmite P/Ca ratio. Because Ca is the major ion which is accounting for stalagmite growth, increases in P/Ca cannot be only due to higher growth rate because a higher growth rate also means a greater accumulation of Ca per unit time.

The incorporation of P in stalagmite calcite may vary due to changes in the pH of dripwater and the evolution of dripwater chemistry during calcite precipitation. Water infiltrating through soil and root zones is acidified by respired CO₂ and dissolves limestone bedrock; higher soil CO₂ leads to lower water pH and greater dissolution. When infiltrating water reaches the lower pCO₂ cave environment, CO₂ begins to degas from the drip water, leading to an increase in pH and the precipitation of CaCO₃. CaCO₃ precipitation can occur on the cave ceiling before the drip water falls to the top of the stalagmite, and this “prior calcite precipitation” affects the pH and concentration of ions in the water forming the stalagmite. This process is simulated by a number of models.

To evaluate the potential effects of prior calcite precipitation and changing drip water pH on P incorporation, we carry out a set of calculations accounting for the change of Ca, pH, dissolved inorganic carbonate (DIC), and dissolved P in drip water. We employ CAVECALC (Owen et al., 2018) to simulate the evolution of the DIC from bedrock dissolution through progressive degassing for a set of 5 initial soil pCO₂ concentrations which lead to a range in initial pH of drip water (Table S1). For each set of conditions, from Cavecalc, we export the sequence of evolution of Ca, DIC, HCO3-, CO3-2, and pH from dissolution to complete degassing when the water is equilibrated with the cave atmosphere. We report the progressive Ca loss using the index fCa, where fCa is the fraction of initial Ca remaining in solution (see Stoll et al., 2023). From a stalagmite sample, one indicator of the soil pCO₂ can be attained from the stalagmite carbon isotope ratio corrected for in-cave fractionation processes (δ13C_init), because δ13C_init becomes more negative at higher soil pCO₂. We use a previous set of Cavecalc simulations (Lechleitner et al., 2021), to provide the δ13C_init value of calcite corresponding to open system dissolution at the given soil pCO₂.

We seek to evaluate the incorporation of orthophosphate in calcite and do not simulate the incorporation of P associated with detrital phases or other accessory minerals. For comparison with the P/Ca, we evaluate also the impact on P/Mg since the drip water concentration of the latter, highly incompatible element is negligibly affected by calcite precipitation. We evaluate each of these 5 CAVECALC runs for four different possible models for the behavior of P incorporation in calcite (Table 2). For the calculations, we define the P/Ca ratio of drip water at initial dissolution (hereafter initial dripwater P/Ca) to reflect the median P/Ca of the host bedrock of La Vallina Cave which is 0.2 mmol mol⁻¹; this ratio is consistent with the modern drip water concentrations in La Vallina being below 0.32 µmolL-1 and could reflect a steady state value where P inputs from rock dissolution are comparable to P loss through groundwaters. We set the initial drip water Mg/Ca ratio at 2 mmol mol⁻¹ for all simulations and employ a DMg of 0.03. Following Dodd et al. (2021) we define P incorporation coefficient (D) relative to Ca, which matches the analytical measurement of P/Ca ratios, although the actual P incorporation mechanism may reflect substitution for the carbonate ion rather than Ca⁺². We calculate the evolution of the drip water P concentration iteratively through the progressive calcite precipitation, evaluating various assumptions for the incorporation of P in calcite. In two models, we specify that the P incorporation is proportional to the solution ratio of HPO4-2 to CO3-2 and therefore evolves during degassing, starting with a maximum D of 1.46 in one scenario and a maximum D of 0.75 in another (Table 2) (D, the partition coefficient, is the ratio of P concentration in calcite to that in drip water). In two other models, we specify a constant P incorporation (D=1, and D=0.5). The partitioning proportional to the solution ratio of HPO4-2 to CO3-2 is consistent with a mechanism of P incorporation dominantly substituting for CO3-2, and is consistent with the observed pH dependence of P incorporation in carbonates (Dodd et al., 2021). A constant D could be indicative of P incorporation as both HPO4-2 and H₂PO4-1, the latter substituting HCO3- (potentially analogous to borate incorporation in calcite). The incorporation of HCO3- as well as CO3-2 in growing calcite remains discussed (Andersson et al., 2016; Huang et al., 2021) hence we include this possibility in our sensitivity analysis. There are limited constraints on the effective P incorporation coefficient in calcite. The experiments of (Dodd et al., 2021) for calcite suggest D=0.47 at pH 8. An additional constraint is provided by the maximum the drip water P concentration (the analytical detection limit of 0.32 µmolL-1 and 61 ppm average Ca concentration (Kost et al., 2023) and the average P/Ca of 0.13 mmol mol⁻¹ for the actively growing stalagmite SNO in La Vallina Cave for which the negligible correlation of P/Ca and Al/Ca suggests limited detrital P contribution (Sliwinski et al., 2023). For this single, rapidly growing stalagmite, we estimate a minimum D of 0.6 but we do not have an estimate for the pH during precipitation of this sample.

To evaluate the changes in snow cover and soil humidity under evolving boundary condition, we employ a previously published simulation (Romé et al., 2022). The simulation features a quasi-idealised glacial climate state with an oscillating AMOC strength, which is triggered by a constant meltwater flux of 0.084 Sv corresponding to the GLAC-1D (Ivanovic et al., 2016) ice sheet history at 17.8 ka BP. When comparing periods of a strong (∼16 Sv) to a weak (∼8 Sv) AMOC state in this simulation, the patterns significant for an AMOC slowdown emerge (i.e., temperature reduction in the North Atlantic realm and Southward shift of ITCZ), allowing us to study the range of plausible conditions above the study site.

The median P/Ca ratio ranges from 0.1 to 0.2 mmol mol⁻¹ across 9 datasets from 7 different stalagmites (GAE and NEI provided data during two distinct periods) (Fig. 2). The PGM-LIG section of GAE features numerous analyses with higher P/Ca than other samples, despite very slow growth and higher signal smoothing than other coeval stalagmites such as BEL and GAR, and features a positive correlation with Al/Ca ratios in the acid soluble fraction (Fig. S3). GAE in this time span also features intervals of anomalously high ²³²Th in the series of dated samples, indicative of intervals with appreciable detrital content (Stoll et al., 2022, 2013). These high P and high Al data points originate from the base of the stalagmite and from an interval of condensed and potentially interrupted growth in the last interglacial (121 to 124 ka BP). For examination of subsequent trends, we evaluate the P/Ca ratios which have Al/Ca ratios <0.04 mmol mol⁻¹, the interval below which Al/Ca and P/Ca exhibit no correlation in GAE. This filter eliminates few points in stalagmites other than GAE (Fig. S3). In recent actively growing stalagmites in Zoolithen cave, detrital contribution to stalagmite P was inferred from high correlations with Al (Riechelmann et al., 2020). This removes singular high P layers such as growth hiatus which may be enhanced by accumulation of (aluminum-bearing) silicates and the impact of microbially dominated carbonate phases following intervals of stalagmite dissolution (Frisia et al., 2012).

We divide the PGM-LIG datasets into three intervals, the last interglacial (127 to 118 kyr BP), the penultimate deglaciation (135 to 127 kyr BP) and the penultimate glacial (145–135 kyr BP). Among samples from the same interval, there are substantial differences in the mean P/Ca ratios, with BEL and GAE exhibiting lower median P/Ca than GAR or NEI (Fig. 3). These differences are not due to contrasting bedrock P/Ca, since the contrasting BEL and GAE are from samples from the same cave. Although the median P/Ca in bedrock from Cueva Rosa is half that of La Vallina, the stalagmite from Cueva Rosa (NEI) does not feature lower mean P/Ca.

Soil pCO₂ and δ13C_init are inversely correlated due to contrasting isotopic compositions of respired CO₂ and atmospheric CO₂, an effect simulated by CAVECALC (Lechleitner et al., 2021).

Most stalagmites spanning PGM to LIG exhibit a wide range in P/Ca for a given δ13C_init (Fig. 7). Models of P incorporation into calcite show that while soil pCO₂ affects drip water pH and may potentially affect P incorporation into stalagmite calcite, this partitioning effect and concomitant evolution of the dripwater through degassing, would give a much smaller range of variation in P/Ca than observed (Figs. 7, S4–S5, Supplement). Stalagmites BEL, GAR, and to a lesser extent GLD, attain peak P/Ca concentrations at intermediate δ13C_init. For this intermediate δ13C_init, the range in P/Ca is >2 fold (up to 4-fold in GAR) larger than could be predicted by variations in fCa according to a range of reasonable models either model 1 (Dmax=1.46) or model 3 (D=0.5) (Fig. 7). At the same time, in GAR, the minimum P/Ca values trend higher with decreasing δ13C_init, consistent with the trend predicted by some models (Figs. 7, S4; Model 2 (Dmax=1.46) for fCa = 0.95) This trend may be attributable to variations in P incorporation. In NEI, samples of moderately high fCa (0.8 to 0.87) closely follow the modeled trend in P/Ca for model 2 (Dmax=1.46), thus some of the variation in P/Ca may be attributable to variations in P incorporation.

A similar comparison of model and data for the GS22 interval also indicates that the range of P/Ca variation exceeds that which is simulated to result from changing P incorporation at a constant dripwater P/Ca (Fig. S6).

The stalagmite records provide evidence for periods of elevated initial drip water P/Ca ratios. Because greater bedrock dissolution would proportionally increase both the P and the Ca (and DIC) concentrations in drip water leaving the P/Ca ratio unchanged, the periods of elevated P/Ca cannot be explained only by greater bedrock dissolution. Rather, they likely reflect increased release of P from non-carbonate reservoirs of P. Independent data on the evolution of vegetation from pollen abundances in coastal sediment cores indicate that there is not a consistent relationship between tree cover and P export. While the abrupt cooling at the onset of GS22 led to decreased vegetation productivity, including a reduction in Atlantic forest pollen from 40 % to <10 % recorded in marine sediment cores off the NW Iberian margin (Goni et al., 2008), during the penultimate deglaciation, temperate tree pollen progressively increased from < 10 % during the PGM to nearly 60 % by 129 ka (Tzedakis et al., 2018) (Fig. 5). Thus, because there is no consistent sign of change in vegetation during the pulses in P/Ca, we suggest that changes in P storage in aboveground living biomass are not the dominant cause of the peaks in P export.

During the PGM to LIG transition, our age models suggest these pulses may have lasted about 1 kyr. Most events fall during the onset of the cold stadial events indicated from rapid positive shift in δ13C_init, but in GS22 the end of the stadial may also feature elevated P/Ca and one event recorded by GAR occurs during a weak stadial in between two cold extremes, while in the different cave setting of NEI the pulses were more linked to stadial recoveries than onsets.

Cold stadial interludes are widely attributed to weakened AMOC and therefore we consider the simulations during weakened AMOC to provide indicative information for how stadial climate in the study region differed from glacial climate conditions. GCM climate model simulations suggest hydrological balance (precipitation-evapotranspiration) in this region was similar during cold events of significantly weakened AMOC as during glacial climate with strong AMOC (Romé et al., 2022). However, the infiltration regime may have changed in response to winter temperature and snowcover. During simulated weak AMOC conditions, monthly average top soil temperatures remained below freezing several months of the year, which does not occur in simulated interstadial simulations, and base soil temperatures are colder (Fig. 8). Furthermore, the winter snowcover is simulated to be significantly higher during the weakened AMOC conditions than during the preceding glacial. In combination with the below freezing temperatures, the increased snowcover may suppress infiltration during winter months of stadial periods and lead to large infiltration peaks during spring snowmelt. Changes in the winter temperature distribution also likely affect the intensity and severity of freeze-thaw cycles, a parameter which could be evaluated in future paleoclimate model studies exporting hourly resolution data from paleoclimate simulations.

Modern monitoring studies suggest several aspects of changing cold season climate which contributed to enhanced leaching of P from soil and biomass. Soil microbial cycling is strongly temperature sensitive and soil microbes store large amounts of inorganic and labile organic phosphorus compounds (Gao et al., 2021). Strong decreases in winter soil temperatures lead to lysis of microbial cells, and a decreased soil microbial population may suppress microbial recycling of P, leading to greater loss to groundwaters. In climate chamber experiments, modern soils from cold, snow rich sites featured an up to 400 % increase in soluble phosphate when subjected to high magnitude freeze-thaw cycles (Kreyling et al., 2020). Across a global dataset, freeze-thaw cycles increased dissolved total P and dissolved inorganic P by 312 % and 115 %, respectively. This effect may partly reflect the P release from the death and lysis of soil microorganisms, but it may also arise because ice crystals disrupt the bonding of soil aggregates and accelerate organic matter release (Gao et al., 2021). Finally, frozen topsoil and accumulation of snow change the hydrological cycle by storing water and preventing infiltration during the period of frozen topsoil. If the rate of melting exceeds the infiltration rate when topsoil remains frozen, spring snowmelt may disproportionately enhance surface runoff, and historically the high rates of nutrient loss during snowmelt have been attributed to overland erosion, especially in agricultural catchments (Costa et al., 2020). When areas formerly continuously covered by winter snowfall are subjected to intermittent melt and rain-on snow events, increased frequency of cold season flushing may accelerate nutrient loss, at least via surface waters (Seybold et al., 2022). But, depending on the infiltrability of the soil, meltwater can also infiltrate into the soil, accelerating nutrient flushing (Granger et al., 1984).

In addition to the direct climatic effects on freeze-thaw frequency and hydrology, the climatic transitions into stadial cold events likely also affected soil pH. More positive δ13C_init values during the stadials indicate reduced soil respiration and lower soil CO₂ concentrations, which would raise the pH of soil water. Thus, the elevated P/Ca at the onset and termination of GS events could also reflect pH-driven mobilization of mineral-bound P, if intermediate pH conditions destabilized a fraction of the soil P pool. At the onset of the stadial, the positive pH shift may have enhanced the dissolution of Ca-phosphate minerals in calcareous soils and reduced P adsorption to Fe- and Al-(oxyhydr)oxides in more acidic soils, thereby increasing the flux of dissolved P to drip waters. This pathway, acting alongside changes in organic P cycling, may therefore also contribute to the magnitude of the observed P/Ca excursions during these climate transitions.Unlike the recent observations of nutrient export or experimental treatment which deal with controls of nutrient release over weeks to months, we have detected events indicating enhanced groundwater P export during hundreds of years to a thousand years. We suggest that during the transitions into stadial cold periods, a combination of more frequent or intense freeze-thaw cycles, suppression of microbial P recycling, and more intense spring infiltration events during snowmelt all increased the P flux in groundwater infiltrating into the cave (Fig. 9). This period of increased flux likely represents a transient imbalance between the P supply from bedrock weathering and the P loss in groundwaters. During several events, during the coldest period of the stadial, P fluxes returned to the lower levels characterizing the period prior to the onset of cold conditions. Potentially, this decline could reflect the depletion of soil P pools following a sustained imbalance in the cycle, given estimates for a mean residence time of P in the soil pool of ≈500 years (Wang et al., 2010). Alternatively or in addition, it may reflect climatic conditions which were less favorable to P export, such as persistent cold with fewer freeze-thaw cycles or soil warming due to the insulation of snowcover which allowed a more effective soil microbial recycling of P. Finally, during some events, enhanced P export also occurs during the climatic recovery from the cold stadial period. Potentially, these transitions may have featured greater frequency of freeze-thaw cycles, or the climatic recovery may have facilitated an increase in bedrock dissolution rates which replenished P stocks sufficiently that P could be exported during these leaching events.