Last glacial millennial-scale hydro-climate and temperature changes in Puerto Rico constrained by speleothem fluid inclusion δO and δH values

We present speleothem fluid inclusion δOf and δHf values from Larga Cave, Puerto Rico, that coverscover the interval between 46.2 to 15.3 ka before present on millennial scale, including the Last Glacial Maximum (LGM) and several stadial and interstadial cycles. The dataset can be divided in two main clusters of stable isotope compositions of the fluid inclusion 20 water with respect to the global meteoric water line which coincide with strong variations in the water content of the stalagmite. In particular, this clustering is found to be climate related, where the firstone cluster comprises samples from cold and dry periods, such as Heinrich and Greenland stadials, as well as parts of the LGM, which exhibit very high δOf and δHf values. We interpret this enrichment as caused by evaporation inside the cave due to enhanced cave ventilation during these colder and drier times. In contrast, in most samples corresponding to warmer and wetter Greenland interstadials, but also for some 25 from Heinrich Stadial 2 and 3, the δOf and δHf values plot on the meteoric water line and modification of fluid inclusion water due to “in-cave” evaporation is found negligible. Consequently, variations of last glacial hydro-climate and temperature in the western tropical Atlantic can be constrained. In general, δOf values from fluid inclusions are up to 3‰ higher than those of modern drip water, which is interpreted as a weaker atmospheric convective activity during the last glacial period. In addition, reconstructed temperatures suggest an 30 average cooling of c. 2-3°C during the LGM compared to modern cave temperature. During Heinrich Stadials 2 and 3, reconstructed cave temperatures yield an additionalaverage cooling of -1.4 ± 2.9 ± 2.68 °C for HS 2 and 4.4 ± 0-3.6 ± 2.2 °C, respectively for HS 3. Higher δOf values of these samples further suggest that the drip water was dominated by orographic rainfall and/or cold fronts, along with weak or even absent convective activity. In contrast, during interstadial phases,

reconstructed temperatures reached nearly modern values, and convective activity was comparable or only slightly weaker 35 than today.

Introduction
Stable isotopes of oxygen (represented by δ 18 O values) are a widely used proxy to deduce information about past climates from speleothems, since the cave drip water reflects the amount-or recharge-weighted annual mean isotopic signature of the rainfall above the cave (Baker et al., 2019;Lachniet, 2009). This value is in turn modulated by several variables, such as 40 temperature, the type and seasonal distribution of precipitation as well as the source moisture trajectory (Lachniet, 2009;Bird et al., 2020;Lases-Hernandez et al., 2020). In the tropics, the isotopic composition of precipitation, δ 18 Op, has been attributed to changes in precipitation amount and convective activity (Dansgaard, 1964;Lachniet and Patterson, 2009;Vieten et al., 2018a). Consequently, tropical speleothem δ 18 O values are usually interpreted in terms of this isotopic 'amount effect' (Medina-Elizalde et al., 2010;Arienzo et al., 2019;Lases-Hernandez et al., 2019). However, in the soil and karst above the cave 45 as well as at the surface of a stalagmite, various (disequilibrium) isotope fractionation effects can obscure the interpretation of stable isotopes in speleothem calcite (Weber et al., 2021;Hansen et al., 2019;Carlson et al., 2020).
In this respect, stable isotopes of oxygen and hydrogen in speleothem fluid inclusions provide a useful proxy, since these represent potentially unchanged aliquots of the original drip water, and thus relicts of past precipitation (Schwarcz et al., 1976;van Breukelen et al., 2008;Affolter et al., 2015). The utility of speleothem fluid inclusion analysis improves if the 50 conditions of carbonate precipitation are well understood, i.e., whether or not the minerals precipitated close to 'equilibrium conditions'. In the case of equilibrium conditions, paleo-temperatures can be reconstructed from combined oxygen isotope ratios of the calcite (δ 18 Oc) and the fluid inclusions, δ 18 Of, based on known temperature dependent isotope fractionation relationships between calcite and water (e.g., Kim and O'Neil (1997); Tremaine et al. (2011); Hansen et al. (2019); Johnston et al. (2013). In continental mid-and high latitude locations, estimations of paleo-temperatures via fluid inclusion δ 2 Hf values 55 alone are feasible, based on the dominant control of temperature on rainfall isotope composition in such settings (Affolter et al., 2019;Demény et al., 2021). This only works, however, if the contribution of the amount effect is negligible. The analysis of speleothem fluid inclusions has been successfully applied worldwide, including tropical locations (Griffiths et al., 2010;Arienzo et al., 2015;Millo et al., 2017;Meckler et al., 2015). Nevertheless, speleothem fluid inclusion records from the Last Glacial Period (c. 115 -11.7 ka) are still sparse, and in many cases, reliable paleo-temperature estimates are challenging 60 (Millo et al., 2017;Demény et al., 2021;Nehme et al., 2020).
Here we present a detailed study of fluid inclusions from a Puerto Rican stalagmite (PR-LA-1), spanning the interval between 46.2 to 15.3 ka. The record encompasses a period which was characterized by rapid climatic fluctuations from cold (stadial) to warmer (interstadial) conditions recurring on millennial time scales (Dansgaard et al., 1984;Peterson et al., 2000). In paleoclimate records from the tropics, the so-called 'Heinrich Stadials' (HS) are clearly represented by dramatic cooling and dry conditions in the western tropical Atlantic (Escobar et al., 2012;Arienzo et al., 2015;Grauel et al., 2016). These conditions were likely caused by a southward shift of the Inter-Tropical Convergence Zone (ITCZ) induced by a near shutdown of the Atlantic Meridional Overturning Circulation (AMOC) (Broccoli et al., 2006;Lynch-Stieglitz et al., 2014;Deplazes et al., 2013;Escobar et al., 2012). Conversely, during the interstadials, the ITCZ was presumably in a more northerly position, accompanied by warmer and wetter climatic conditions and higher SSTs in the (sub-) tropical Atlantic (Schmidt et al., 70 2006a;Peterson et al., 2000;Sachs and Lehman, 1999).
The speleothems PR-LA-1 was selected because its stable isotope values (δ 18 Oc and δ 13 Cc) and Mg/Ca ratios demonstrated a pronounced response to local hydro-climate on the millennial-scale (Warken et al., 2020). To further elucidate the millennialscale climate changes during the last glacial stadials and interstadials, we thus conducted a large number of fluid inclusion measurements following prominent δ 18 Oc changes in stalagmite PR-LA-1. 75

Site description
Larga Cave (18°19'N 66°48'W, Fig. 1) is located 350 m above sea level in the north central karst region of Puerto Rico. The 90 overburden of Larga Cave is comprised of the Oligocene Lares limestone Fm., which is covered by a thin layer of soil. The host rock is exposed at some places, and vegetation consists today of dense tropical forest. The main passage in the cave is almost horizontal and has a total length of 1400 m with a ceiling height of up to 30 m (Miller, 2010). Stalagmite PR-LA-1 has a total length of 185 cm and was found lying on the cave floor in the main passage, where the rock overburden is 40 to 80 m thick. The stalagmite was deposited between 15.4 to 46.2 ka before present with a hiatus between 35.5 and 41.1 ka (Fig. 1,  95 Warken et al. (2020) (Vieten et al., 2016). This seasonal pattern results from a temperature-driven, very well-ventilated cave environment in winter and low ventilation in summer. During the winter season, a high buoyancy contrast at night leads to maximal cave ventilation, while it is at a minimum during summer, when cave temperatures are lower than in the outside environment. RH is mostly close to 100%, but has been found to decrease to values to 90-95% for several weeks when ventilation intensifies in 105 winter months (Vieten et al., 2016;2018a).
As typical for Caribbean locations, the precipitation regime over Larga Cave is characterized by a drier winter season, from December to April and a much wetter season during the rest of the year. During the drier season, occasional light showers occur imbedded in trade winds, which are interrupted by occasional cold fronts emanating from North America (Scholl and Murphy, 2014). In contrast, during the remainder of the year convective rain is formed by easterly moving low pressure systems 110 which occasionally develop into tropical depressions accounting for the major part of total precipitation (Scholl and Murphy, 2014). On inter-annual time scales, the δ 18 O values of rainfall in Puerto Rico show a negative correlation with precipitation suggesting an isotopic 'amount effect' of c. -0.1 ‰ per 100 mm of rainfall amount (Govender et al., 2013;Scholl et al., 2014;Warken et al., 2020). Over the course of monthly drip water monitoring (2013 -2017), the drip water composition at sites with a maximum distance of 10m to the location where stalagmite PR-LA-1 was found shows only little seasonal variation 115 with a mean value of δ 18 Ow = −2.6 ± 0.3‰, and δ 2 Hw = −9.5 ± 0.2‰ (VSMOW). The seasonal signal in the rain water is strongly smoothed by the soil and epikarst above the cave with a mixing time of at least several months to years, suggesting that Larga speleothems time resolution record is at least inter-annual.

Fluid inclusion stable isotope composition
Solid samples for fluid inclusion isotope analysis were cut using a band saw along the growth axis of stalagmite PR-LA-1.
High growth rates of this specimen of up to 1mm per year allow the sampling of distinct millennial scale climate oscillations 125 evident in the δ 18 Oc record, such as D/O or Heinrich events. Replicate samples were taken along the corresponding growth layers to minimize the effect of integrating over too long time intervals. Individual sample sizes were between 0.21 -1.34 g.
The speleothem fragments were placed in a copper tube and installed in a preparation line connected to cavity ring down spectrometer (CRDS, L2130-I, Picarro) at the Institute of Environmental Physics, Heidelberg University (Weißbach, 2020).
The samples were subsequently crushed using a hydraulic press. To avoid theAfter connecting the sample, the setup requires 130 approximately 90 minutes until the water vapour background has reached stable conditions. When the standard deviation of the water vapour concentration is below 20 ppmV over 30 minutes, the sample is crushed with a hydraulic crusher. To avoid a CRDS memory effect, a stable water vapour background was generated to ensure constant saturation of the cavity without condensation (Affolter et al., 2014;De Graaf et al., 2020). This was achieved by mixing dry nitrogen as a carrier gas with water of known isotopic composition, in awhile the line, including the tube that holds the sample, was held at a constant temperature 135 of 120°C.
The mixture of background water and sample signal is measured in the CRDS analyser, and evaluated similar to the method of Affolter et al. (2014) with a Python code. The isotopic composition and released water volume is calibrated with in-house water standards and glass capillaries with known volumes in the µl range, respectively. These are inserted into the copper tube and connected to the preparation line, similarly as the speleothem samples. The in-house water standards are analysed regularly 140 against the international primary reference material VSMOW2 and SLAP2. The long-term precision of single stable isotope measurements is 0.5 ‰ for δ 18 O and 1.5 ‰ for δ 2 H, while for released water volumes above 1.0 μl, the precision improves to 0.2‰ for δ 18 O and 0.4‰ for δ 2 H (Weißbach, 2020). In case of water volumes < 0.2 μl, the result of a single measurement may become less precise and is therefore reproduced by replicate measurements to verify the result. If not stated differently, the stable isotope composition of fluid inclusions is given relative to Vienna Standard Mean Ocean Water (VSMOW). 145

Water content and stable isotope composition of speleothem fluid inclusions
In total, 64 fluid inclusion water isotope measurements were performed from 30 different depth intervals. The results of the single measurements are shown in Table S1, while the mean values of the replicate measurements are given in Table S2. The 160 mean water content of PR-LA-1 calcite for the different depth intervals shows a variation from close to 0 to 3 μl/g (Fig. 2).
Sections with a relatively high water content, e.g., around c. 2824 or 34ka28 ka, also show a high (or at least highly variable) growth rate, while parts with a comparably low water content, such as at c. 20, 26, or 32ka are mainly in sections with relatively slow growth rates. The mean fluid inclusion water δ 18 Of values vary between -1.9 and +11.4‰, and mean δ 2 Hf values are between +6.9 and +44.9‰ (VSMOW) (Fig. 3a). The data exhibit a systematic 18 O and 2 H enrichment compared to modern 165 drip water stable isotope values. This general pattern is also reflected in the slope of the linear regression line fitted to the data, which is with 2.42 ± 0.25 significantly lower than expected for a LMWL. This slope agrees with those of globally distributed evaporation lines for lakes and soil waters (Gibson et al., 2008) as well as evaporation experiments (Hu et al., 2009).
Furthermore, the data can be allocated to two different clusters, depending on the measured water content of the samples. The first cluster comprises samples which are mainly located on or scatter around the GMWL. Most of these are also characterized 170 by a water content greater than 0.55 µl/g. In contrast, a lot of samples show more positive δ 18 O and δ 2 H values and clearly plot off the GMWL along the linear regression line. In contrast to the first set of samples, most of these have a water content lower than 0.55 µl/g. Only one sample clearly does not fit into this pattern and plots away from the GMWL despite a water content >0.55µl/g (sample 32b), which therefore are also allocated to cluster (2). Conversely, two samples plot close to the MWLs despite a relatively low water content (samples 5 and 8). The relationship between fluid inclusion stable isotope values and 175 water content is also expressed numerically with a negative correlation of δ 18 Of and the measured water content with r = -0.72 (r² = 0.51) for logarithmic fitting (supplementary Figure S 2).

Calculation of paleo-temperatures using the oxygen isotope equilibrium thermometer
In order to assess the applicability of the oxygen isotope equilibrium thermometer based on the mean fluid inclusion δ 18 Of values, the average carbonate δ 18 Oc values were calculated from the intervals covered by the depth range of the measured fluid 180 inclusion samples, respectively (Table S2). Figure 4a shows that the previously mentioned pattern with two clusters of samples is also reproduced when comparing the fluid inclusion δ 18 Of with the carbonate δ 18 Oc values. Samples with a high water content, which plot on the GMWL in Fig. 3a, are clearly separated by the samples with a lower water content. While cluster (1) samples with a high water content show a positive correlation between δ 18 Oc and δ 18 Of values, cluster (2) samples do not follow the same trend. Similarly, the fractionation between speleothem calcite and fluid inclusions, 18 εc/f = 1000·lnα (with α = 185 (1000 + δ 18 Oc,)/(1000 + δ 18 Of), and δ 18 Oc and δ 18 Of in VSMOW) shows a clear dependence on the δ 18 Of value for the evaporated samples of cluster (2) (Fig. 4b), but no relationship with carbonate δ 18 Oc values (Fig. 4c). In contrast, for the cluster (1) samples, no relationship between 18 εc/f and δ 18 Of values is observed (Fig. 4b). Figure 4 also suggests that fluid inclusion samples 9b, 32b, 27, 32, and 36 are also enriched in 18 O, and have to be allocated to cluster (2). between Heinrich stadials, Greenland stadials, and Greenland interstadials. Samples analysed from the LGM are also subdivided in presumably more warm/humid and relatively cool/dry periods. Numbers refer to the sample IDs as shown in Tables S1 and S2.  (Table S2). The uncertainty of the calculated fractionation is estimated from propagating the analytical error of the fluid inclusion measurements, as well as the range of the δ 18 Oc calcite value due to dating and sampling uncertainty. Blue symbols indicate samples with a water content >0.55 µl/g, and light blue symbols a water content <0.55 µl/g. The ellipses indicate the samples which are allocated to cluster (2) due to large fractionation effects in δ 18 Of. The grey lines indicate the water-calcite oxygen isotope fractionation relationship after Tremaine et al. (2011) for cave temperatures of 15, 20 and 25°C, respectively. The ellipses indicate the samples which are allocated to cluster (2) due to large fractionation effects in δ 18 Of. All these samples also plot off the water-calcite oxygen isotope fractionation curve for 25°C.
For a first-order temperature calculation from 18 εc/f , three different isotope fractionation relationships have been tested (Kim 215 and O'Neil, 1997;Tremaine et al., 2011;Johnston et al., 2013). All results are listed in Table S2, where the given uncertainties of the paleo-temperatures result from propagating the analytical error of the fluid inclusion measurements, as well as the range of the δ 18 Oc calcite value due to dating and sampling uncertainty. All paleo-temperatures reported and discussed in the main text are calculated using the JohnstonTremaine et al. (20132011) equation, which has been developed from cave monitoring and modern speleothem calcite and yields realistic temperatures compared to the modern cave air temperature (Table S2). Note 220 that we discard samples from the temperature discussion where replicate measurements exhibit a poor reproducibility in terms of largely variable water content and large isotopic differences, in order to focus on the most robust paleo-climatic inferences (Table S2). Note also that we diddo not calculatereport temperatures for the samples of cluster (2), which did not contain sufficient water and clearly show significant influence from fractionation through evaporation (Table S2). We regard the temperature results obtained on the 12 replicate samples from cluster (1) as reliable within the large uncertainty of 1.5 -2.5°C. 225 The calculated temperatures scatter around the modern cave temperature of 22.8°C, with paleo-temperatures varying between 15.216 ± 1.78 °C (sample 8, 21.39 ± 0.25 ka) and 24.025.7 ± 1.24 °C (sample 24, 33.52 ± 0.52 ka), and a mean value of 20.021.1 ± 2.57 °C.

Categorization of the fluid inclusion isotope data
For the fluid inclusion isotope data from cluster (1), δ 18 Of and δ 2 Hf values scatter around the LMWL (Govender et al., 2013;Putman et al., 2019;Scholl and Murphy, 2014;Scholl et al., 2014). Thus, we suggest that these samples represent pristine fossil drip water. The second cluster of fluid inclusion isotope values contains mainly samples with water contents below 0.55µl/g, which all show increased δ 18 Of values which are noticeably more fractionated than hydrogen isotopes. In addition to 235 the categorization with respect to water content, the fluid inclusion samples can be further associated with stadial and interstadial phases during the Last Glacial according to their stable isotope and elemental signatures in speleothem calcite (Fig.   3b). The fluid inclusion samples representing warm Greenland Interstadial (GI, red triangle) plot close to the MWL and can be all associated with cluster (1). In contrast, the samples formed during the colder intervals (HS; GS, LGM) are mainly attributed to cluster (2) and trend away from the MWL. In addition, for Last Glacial Maximum (LGM) samples, the 240 differentiation between the two clusters appears to be valid, with relatively warm/humid phases plotting closer to the MWL, and relatively dry/cool phases rather clustering with the GS and HS samples (Fig. 3b). This demonstrates that our dataset allows us to distinguish cooler and/or drier climatic background states during Heinrich and Greenland Stadials, represented by cluster (2) samples, from the presumably warmer and more humid phases during Greenland Interstadials, which comprise cluster (1). Note that evaporation driven fractionation must indicate a general annual decrease of humidity in the cave spanning 245 over decades required to trap the water in the calcite. Therefore, we suggest that samples of cluster (2) reflect periods with stronger cave ventilation as recorded today in the thermal contrast of cave and ambient air in winter. This categorisation suggests that the underlying process, which causes this clustering and affects the fluid inclusion isotopic signatures, is purely climate related.
Previous studies haveIsotope effects due to partial evaporation of fluid inclusion water can in principle also occur during the 250 warm-up phase in the crusher when fluid inclusions are insufficiently sealed to withstand the pressure that builds up during warm-up. Based on our experience, water loss in the warm-up phase causes noticeable disruptions in the water background of the analyses, but this has been observed mostly for diagenetically compromised samples so far. We note that the water vapour background was continuously monitored before crushing, and no anomalies were observed for the samples from PR-LA-1 (Weißbach, 2020). Figure S 1 also shows no clear relationship between shape and distribution of fluid inclusions in PR-LA-1 255 nor do we see signs of diagenetic alteration in stalagmite PR-LA-1, which could have favoured inclusion leaking during heating. Previous studies have also invoked recrystallization-induced oxygen isotope changes in inclusion-hosted water of speleothems (Demény et al., 2016). This process is unlikely to be systematically related with climate and would shift the fluid inclusion δ 18 O values towards more negative values, which is opposite to the observed pattern in this study. Moreover, no systematic signs of diagenetic alteration have been observed for stalagmite PR-LA-1 (Warken et al., 2020). 260 Other studies suggest that in situ oxygen isotope exchange of fluid inclusion water with the host carbonate is unlikely to cause significant isotope fractionation (Uemura et al., 2020;Demény et al., 2021). The systematic enrichment of cluster (2) samples indicates that the stable isotope composition of the low water content samples has been affected by fractionation processes in the gas phase as 18 O is enriched over H isotopes, just as in non-equilibrium evaporation of thin water films (Gat et al., 1994).
Evaporation is thus aThus, all information combined suggests evaporation inside the cave on the surface of the stalagmite as 265 the most likely scenario causing the lower slope of the water isotope relationship of 2.42 ± 0.25 (Fig. 3) via a temperature and humidity related climatechange (Gibson et al., 2008;Hu et al., 2009).

Climate-related evaporation effects on fluid inclusion isotope values in Larga cave
Recent studies have suggested that in-cave evaporation may affect the isotopic composition of speleothem fluid inclusions (Nehme et al., 2020;Weißbach, 2020). A potential climate-related driver of this evaporative fractionation that has not been 270 described previously is temperature-driven ventilation. This effect has been observed in Larga Cave today, especially during   (2), 18 εc/f is largely independent from δ 18 Oc values (Fig. 4c), while a very strong correlation with δ 18 Of values is evident (Fig. 4b). This argues for 285 a significant influence of evaporation, which would have a strong effect on the fluid inclusion water, but only a minor effect on the precipitated carbonate. The reason for the different effect of evaporation on the δ 18 Of values of the fluid inclusion water and the precipitated calcite is related to the very different time-scales of precipitation and oxygen isotope exchange between water and dissolved bicarbonate. Precipitation of calcite occurs on a time scale of a few hundred seconds (Dreybrodt and Scholz, 2011), dependingamongst other parameters -on cave temperature and film thickness. The time-scale of oxygen 290 isotope exchange between water and dissolved bicarbonate, in contrast, is in the range of hours to days (Beck et al., 2005;Dreybrodt and Scholz, 2011), and thus orders of magnitude longer.
When the drip water enters the cave, it is reasonable to assume that the δ 18 O value of the dissolved bicarbonate is in oxygen isotope equilibrium with the δ 18 O value of the drip water. When the drip water then falls onto the surface of a stalagmite, degassing of excess CO2 occurs within a few seconds (Dreybrodt and Scholz, 2011) with a negligible effect on the δ 18 O value 295 of the dissolved bicarbonate. Then, precipitation of CaCO3 starts immediately, and the δ 18 O value of the bicarbonate that was previously equilibrated with the drip water is imprinted on the δ 18 O value of the precipitated calcite. We note that oxygen isotope fractionation during precipitation of speleothem calcite is a complex process (e.g., Dreybrodt and Scholz, 2011;Guo and Zhou, 2019;Hansen et al., 2019;Scholz et al., 2009), and oxygen isotope fractionation does in most cases not occur under conditions of oxygen isotope equilibrium. DuringAfter rapid precipitation of speleothem calcite, the drip water remaining on 300 the surface of the speleothem, which forms the fluid inclusion water when it is eventually trapped, may beis progressively affected by evaporation, which would leadleads to increasing δ 18 O and δ 2 H values of the water., and finally to fluid inclusions that are not in isotope equilibrium with the host-calcite. However, due to the very long time-constant of oxygen isotope exchange between water and calcitebicarbonate, it takes hours to imprint the oxygen isotope signature of the (evaporated) water on the dissolved bicarbonate.HCO3 -, and hence, the precipitated calcite. In addition, since the Ca concentration of the 305 drip water drops exponentially, only a minor component of calcite could still be precipitated from the remaining solution (Scholz et al., 2009;Mühlinghaus et al., 2007;Deininger et al., 2012;Dreybrodt and Scholz, 2011) Ventilation in Larga Cave is mainly driven by temperature differences between the atmosphere and the cave (Vieten et al., 2016). During the dry winter season, over the course of the day, the part of the cave hosting the studied speleothem is warmer 315 than the atmosphere inducing a buoyancy-driven air flow. According to our results, evaporation effects are strongest during the comparably cooler and drier periods of the record. This in turn suggests that cave ventilation was particularly pronounced in these times, which would argue for a stronger seasonal temperature contrast during stadial intervals. This observation seems to beis consistent with the hypothesis of enhanced northern-hemisphere seasonality during abrupt cooling events (Denton et al., 2005). Similarly, Ziegler et al. (2008) reported persistently high summer temperatures of the Atlantic warm pool also 320 during north Atlantic cold events, such as Heinrich stadials, suggesting that these seem to influence mainly the winter climate conditions in the Caribbean. It is therefore plausible that a pronounced and persistently strong seasonal temperature difference with particularly cool temperatures during the winter season enhanced winter type ventilation conditions in Larga cave. Hence, ventilation, and thus, the occurrence of evaporation on the stalagmite's surface would have been significant.

Applicability of the oxygen isotope equilibrium thermometer 325
The previously discussed processes have implications for the applicability of the oxygen isotope equilibrium thermometer using fluid inclusion and carbonate δ 18 O values from PR-LA-1. Post-depositional (but pre-entrapment) enrichment of 18 O in fluid inclusion water due to evaporation apparently results in high 18 εc/f fractionation values between fluid inclusion water and calcite, because such influenced samples do not preserve the original δ 18 Of signal. Samples plotting off the MWL are thus discarded from further inferences. Note, that evaporation effects would increase reconstructed temperatures by a factor of two 330 to even three, hence, temperature reconstruction is very sensitive to evaporation, which must be considered when interpreting meaningful temperature values as likely being maximum temperature estimates. Figure 4 shows that also samples, which plot on the GMWL in Fig. 3, might show signs of evaporation. This is, e.g., the case for the sample taken from Heinrich Stadial (HS) 1 (15.9 ka). Here, other paleoclimate records as well as the PR-LA-1 calcite δ 18 O values and elemental proxies suggest that this phase was the coolest and driest in Puerto Rico and the Caribbean realm (Warken et al., 2020;Escobar et al., 335 2012;Grauel et al., 2016). The temperature calculated from the PR-LA-1 fluid inclusion value, however, would be considerably warmer than at present-day. As outlined above, we have to assume that evaporative effects also influenced speleothem and fluid inclusion formation during HS 1, shifting the δ 18 Of to heavier (more positive) values. This demonstrates that samples may be influenced from evaporation despite a position on or close to the MWL. We therefore consider all calculated temperatures as a maximum estimate, and results lower than +3°C compared to modern temperatures as in first order plausible. 340

Fluid inclusion δ 18 O values suggest reduced convective activity during the last glacial
Despite their position close to the MWL, δ 18 Of values in cluster (1) are significantly higher by +2.6 ± 1.3 ‰ compared to modern drip water δ 18 Ow. At the time of deposition of speleothem PR-LA-1 (46.2 to 15.3 ka), the δ 18 O signature of the (sub)tropical Atlantic surface water was higher by about 1.5‰ due to larger global continental ice volume and increased 345 salinity of the western tropical Atlantic (Hagen and Keigwin, 2002;Lea et al., 2003;Schmidt and Spero, 2011). Consequently, glacial drip water represented by fluid inclusion δ 18 Of values must be corrected for this 1.5‰ offset from present-day. Millo et al. (2017) reported a similar enrichment of about 4‰ compared to modern drip water in last glacial fluid inclusion water in a subtropical stalagmite from Brazil, and a similar glacial-interglacial offset has although been observed in δ 18 O values of stalagmites from Guatemala (Winter et al., 2020) and Florida (van Beynen et al., 2017). 350 We interpret the offset in cluster (1) to reflect a generally weaker convective activity during the last glacial period compared to modern conditions and associated reduced summer rainfall (e.g., Arienzo et al., 2019;Winter et al., 2020;Lases-Hernandez et al., 2019;2020;Bird et al., 2020). Today, Puerto Rico rainfall is isotopically enriched when it is formed by occasional northerly cold fronts as well as orographic processes of trade winds (Scholl and Murphy, 2014). In contrast, isotopically lighter precipitation is associated with convective rain, such as hurricanes and tropical storm events. The type of precipitation, i.e., 355 convective vs. orographic, depends to some extent on the meridional location of the ITCZ as well as related Caribbean SSTs (Winter et al., 2020). In this study, the mean of the reconstructed fluid inclusion temperatures for the LGM (c. 25 -18ka) is 19.420.6 ± 3.32 °C. Since our reconstructed temperatures can be regarded as upper limits, due to an underlying imperceptible amount of evaporation, this supports generally cooler temperatures than today over the course of the LGM. The observation of a glacial-interglacial temperature difference of at least 2-3°C highlights the results of previous Caribbean SST records 360 (Rama-Corredor et al., 2015;Lea et al., 2003;Schmidt et al., 2006b). Such low temperatures were presumably too cold to sustain strong convective activity during the LGM. Accordingly, Winter et al. (2020) suggested that deep atmospheric convection during the Holocene was triggered by exceeding a SST threshold of 27.5°C in the tropical ocean leading to a shift of the hydro-climatic regime from a drier state and weak convection during the last glacial to a humid state and strong convection after 9 ka over Central America. 365 The mean value of Puerto Rican fluid inclusion temperatures obtained during MIS 3 (45.5 -29 ka) is 20.421.7 ± 2.6 °C, which is on average 1°C warmer than the LGM.. Despite the relatively large uncertainties, this is consistent withmay support a latewarmer MIS 3 warmer than the LGM, which has been found in the Guiana Basin (Rama-Corredor et al., 2015) and Brazil (Millo et al., 2017). For Central Mexico, Caballero et al. (2019) report a mean annual temperature of MIS 3 comparable to the present-day. From a global perspective, MIS 3 was c. 2°C warmer than the LGM, due to a northern hemispheric summer 370 insolation maximum and resulting smaller ice sheets (Van Meerbeeck et al., 2009;Svendsen et al., 2004). According to Rama-Corredor et al. (2015), western tropical Atlantic SSTs were close to the threshold of 27.5°C during certain intervals of MIS 3 ( Fig. 5), allowing sustainment of convective activity at least temporarily and/or over a longer time interval over the course of the year. Consequently, relatively wet environmental conditions -even though still drier than those of the Holoceneare supported by records from Mexico, Guadeloupe and the Bahamas (Lozano-García et al., 2015;Royer et al., 2017;Arienzo et 375 al., 2017).

4.2.2
Millennial-scale climate variability in the western tropical Atlantic On the millennial scale, fluid inclusion δ 18 Of (and δ 2 Hf) values exert pronounced swings between very high values of up to 390 2‰ (30‰) during stadial phases, and down to 0‰ (15‰) during interstadial phases between 45 and 27 ka. The pattern largely follows the evolution of the high-resolution calcite stable isotope record (Fig. 2). Vertical blue bars in Fig. 2 (Fig. 5, Table S2). 395 Accordingly, we estimate a mean,an minimum average cooling compared to the present cave air temperature of -1.4 ± 2.9 ± 2.68 °C for HS 2 and -4.43.6 ± 2.2 °C for HS 3. These values are consistent with the results of Arienzo et al. (2015), who found an average temperature decrease of 4°C across Heinrich stadials 1 to 3 for the Bahamas. ManyIn addition, many studies document that HS 1 exhibited the most pronounced cooling of all Heinrich stadials (Arienzo et al., 2015;Grauel et al., 2016).
HS1 was also more pronounced in the isotopic and elemental calcite signals of our sample compared to HS2 or HS3 (Warken 400 et al., 2020). Escobar et al. (2012), and Grauel et al. (2016) found an even stronger cooling down to 10°C for HS1 in Lake Petén Itzá, Guatemala (Fig. 5). We note, however, that these studies used bottom dwelling ostracods, which may have recorded winterbiased lake temperatures. In addition, the strong maritime climate of Puerto Rico located in the eastern Caribbean makes less pronounced temperature variations appear reasonable. In addition to absolute temperature estimates, the samples which show signs of evaporation from cluster (2) suggest that during these times an increased seasonal temperature difference 405 between persistently warm summer temperatures but cooler winter temperatures lead to enhanced cave ventilation. The timing of these samples which includes not only the Heinrich stadials but also Greenland stadials and certain intervals of the LGM is highlighted in Fig. 5. This further illustrates that our dataset suggests that enhanced seasonality may have occurred during these intervals, which would be indicative of a persistent summer expansion of the Atlantic warm pool, but relatively cool winter temperatures. 410 Besides the temperature drops associated with Heinrich stadials, the expression of Greenland interstadials is still not well constrained in the American tropics (Warken et al., 2019). The mean temperature calculated from samples associated with interstadial events in speleothem PR-LA-1 is 22.08 ± 1.89°C (samples 15, 24, 28, 35), which is in the range of present-day cave temperature, and on average 1.6°C warmer than the MIS 3 mean temperature.. In comparison to the average stadial temperatures of 18.6 ± 1.119.7 ± 0.9°C (samples 16, 18, 19), this indicates a difference in the range of up to 3.4 ± 2.8°C 415 between stadial and interstadial phases in Puerto Rico. In the tropical Guiana Basin at 7°N latitude, abrupt glacial SST variations on the millennial scale are in the range of about ±1°C (Rama-Corredor et al., 2015). Further north, Them Il et al.
(2015) report a shift of c. 2°C associated with stadial/interstadial transitions in the Florida Strait (24°N), whereas shifts by 2 to 5°C have been observed in subtropical SSTs, such as at the onsets of GI8 or 12 at the Bermuda Rise at 33°N (Sachs and Lehman, 1999). The available temperatureDespite the large uncertainties associated with the oxygen isotope equilibrium 420 thermometer, our results are thus consistent with these previous reconstructions thus form a north-south transect and suggest a progressive attenuation of millennial-scale temperature oscillations towards the equator.
Our results thus show that also on Puerto Rico the interstadial phases in the western tropical Atlantic were associated with relatively warm temperatures, which may have been even comparable to present-day. Such warm temperatures favoured 425 convective activity, and thus, higher precipitation amounts in the region. Marine reconstructions showed that interstadials were associated with reduced upper water column salinity, while stadial events were associated with increased salinity (Them Il et al., 2015). These observations support our results that meridional shifts in the position of the ITCZ in response to changes in the strength of the AMOC affected the evaporation-precipitation ratio in the western tropical Atlantic (Schmidt and Spero, 2011;Arienzo et al., 2017;Peterson et al., 2000;Schmidt et al., 2006a;Warken et al., 2020). 430

Conclusions
We demonstrate that the stable isotopic composition of fluid inclusions can be strongly influenced by evaporative effects. In Larga Cave, the temperature driven ventilation potentially favoured such conditions, and our data suggest that this effect may have been especially pronounced during cold and dry conditions, such as during Heinrich and Greenland stadial phases and parts of the LGM. During warmer Greenland interstadials, most samples suggest that isotopic modification due to evaporation 435 is negligible. In Larga cave, this process is apparently climate-related, making it possible to omit the potentially affected subsamples for paleo-temperature calculation.
Our dataset thus provides valuable information about last glacial hydro-climate and temperature variability in the western tropical Atlantic and allows to constrain climatic changes on the millennial to orbital scale. For the presumably coldest phases of the last glacial, the Heinrich stadials, we observe a minimum cooling of -1.4 ± 2.9 ± 2.68 °C during HS 2 and of 4.4 ± 0-3.6 440 ± 2.2 °C during HS 3. In contrast, during interstadial phases, temperatures were on average 22.023.4 ± 1.87°C , which is comparable to modern values. The analysis of stable water isotope values in speleothem fluid inclusions further suggests that, compared to present-day, the precipitation regime in the western tropical Atlantic was characterized by generally weaker convective activity during the glacial period as compared to present-day. Especially during the LGM, average temperatures were at least 2-3°C cooler, and thus too cold to sustain deep atmospheric convection. The intermediate background climate 445 state during MIS 3, with average temperatures about 1°C was potentially slightly warmer than the LGM, and favoured an increased sensitivity of western tropical Atlantic hydro-climate to perturbations of the AMOC and northern hemispheric temperature changes on the millennial scales. In this period, the region experienced significant shifts between a slightly weaker but still active -convective regime during the warmer interstadial phases and a dry/non-convective regime during the colder stadial phases. 450