The PALMOD 130k marine palaeoclimate data synthesis version 2

Palaeoclimate data hold the unique promise of providing a long-term perspective on climate change and as such can serve as an important benchmark for climate models. However, palaeoclimate data have generally been archived with insufficient standardisation and metadata to allow for transparent and consistent uncertainty assessment in an automated way. Thanks to improved computation capacity, transient palaeoclimate simulations are now possible, calling for data products containing multi-parameter time series rather than information on a single parameter for a single time slice. To confront transient simulations that span the last glacial-interglacial cycle with palaeoclimate data, we have compiled a multi-parameter marine palaeoclimate data synthesis that contains time series spanning 0 to 130 000 years ago. Jonkers et al. (2020) published the first version of the PALMOD 130k marine palaeoclimate data synthesis and described the data synthesis strategy and the contents and format of the data product in detail. Here we present a major update of the data product that markedly increases both the spatial and temporal coverage. Version 2 of the synthesis contains 2286 time series of eight palaeoclimate parameters from 475 individual sites, each associated with rich metadata, age–depth model ensembles, and information to refine and update the chronologies. It contains 468 time series of benthic foraminifera δ¹⁸O; 357 of benthic foraminifera δ¹³C; 423 of near sea surface temperature; 482 and 273 of planktonic foraminifera δ¹⁸O and δ¹³C; and 128, 111 and 44 of carbonate, organic carbon and biogenic silica content, respectively. Compared to version 1, all radiocarbon ages have been recalibrated and the age-depth models updated. In addition, near sea surface temperature estimates based on planktonic foraminifera $\mathrm{Mg}/\mathrm{Ca}$ and on $U_{\mathrm{37}}^{K^{\prime }}$ have been recalculated using a single calibration thus ensuring global comparability and comprehensive assessment of their uncertainty. The data product is available in two formats (R and LiPD) facilitating use across different software and operating systems and can be downloaded at https://doi.org/10.1594/PANGAEA.984602 (Jonkers et al., 2026). This data descriptor presents our updating methodology and describes the contents and format of the data product in detail and concludes with recommendations on palaeodata stewardship to increase the reusability of such data.

1 Introduction

This article describes an update of the PALMOD 130k multiproxy marine palaeoclimate data synthesis (Jonkers et al., 2020). Palaeoclimate data hold the unique promise to inform about climate states and climate variability on time scales beyond the instrumental period. As such, they may help to characterise natural climate variability, shed light on the mechanisms of climate change and provide analogues for future climate; all crucial insights in a rapidly changing world. Marine sediments and the compounds and fossils contained in them allow the reconstructions of largely continuous, long records of past ocean physical and chemical conditions. Consequently, the multiproxy, global-scale PALMOD 130k synthesis provides a basis to evaluate the importance of the ocean for regulating Earth's climate and biogeochemical cycles on a global scale.

The synthesis contains palaeoceanographic and sedimentological data from sediment cores that cover the last glacial–interglacial cycle (up to approximately 130 000 years ago). During this time Earth has undergone marked changes in climate that involve the atmosphere, continental ice sheets and the ocean (Jonkers et al., 2023; NGRIP Members, 2004; Waelbroeck et al., 2002). Up to the influence of humans on global climate, which accelerated around 1950 CE (Steffen et al., 2015), these changes were ultimately controlled by variations in the Earth's orbit around the Sun (Hays et al., 1976). This orbital forcing triggered a multitude of (non-linear) feedback mechanisms causing global and regional climate to vary much faster than the forcing. A large amount of palaeoclimate data that span this period has been collected over the past decades, rendering the last glacial–interglacial cycle well-suited to study the natural variability of ocean and climate conditions across timescales and in particular the interaction between the cryosphere and the rest of the climate system. This is especially relevant given the projected, but uncertain, (partial) disappearance of continental ice-sheets with ongoing anthropogenic warming. The high-density data availability also makes the period an excellent test-bed for climate model simulations.

Whilst for a long time, palaeoclimate modelling focussed on the simulation of timeslices, increasing computational power now also enables long transient simulations. For instance, the paleoclimate modelling intercomparison project (PMIP) now also provides protocols for the simulation of the last deglaciation (Ivanovic et al., 2016) and several simulations spanning this period exist (Liu et al., 2009; Mikolajewicz et al., 2025). This development allows investigating climate dynamics in a new way, but also puts new demands on the (palaeo)observational data needed for model benchmarking. Specifically, paleoclimate time series need to be associated with comprehensive and consistent estimates of their (chronological) uncertainty. Moreover, the stochastic aspects of climate also call for new ways of comparing transient simulations with observations. For instance, given the uncertainty about whether suborbital abrupt climate variability is forced or unforced, the observed temporal evolution of climate may represent just one of many possible trajectories. Therefore, data-model comparison strategies that are independent of the exact timing of events need to be considered (Weitzel et al., 2024).

A particular strength of paleoceanographic data is their multiproxy nature. We can hence gain a more comprehensive insight into the climate system by studying it from a multi-proxy, or multi-parameter, perspective. Moreover, a multi-parameter approach is also likely to be more diagnostic for model performance than a single climate parameter approach (Kurahashi-Nakamura et al., 2017). However, at present, multi-parameter approaches are difficult because different parameters from the same archive (sediment core) have often been generated in different studies, and the current poor state of data archiving poses serious challenges for reuse. These challenges arise mainly from the non-standardised and fragmented way palaeoceanographic data (and palaeoclimate data in general) are archived and the scarcity of machine-readable metadata included in datafiles. Automated and reproducible analyses of palaeoceanographic data thus remain time consuming, even though efforts are underway to alleviate these problems and increase the ease of data reuse (Emile-Geay and Eshleman, 2013; Morrill et al., 2021). Still, such approaches require standardisation and synthesis of raw and inferred data and of metadata from various sources.

The PALMOD 130k marine palaeoclimate data synthesis is an attempt to overcome the fragmented archiving of paleoceanographic data sets and contains multiple climate-relevant parameters in a single framework. The synthesis combines data on sediment composition (carbonate, total organic carbon and opal content), planktonic and benthic foraminifera stable oxygen and carbon isotope ratios and estimates of seawater temperature based on various proxies. These parameters were selected following discussions within the PALMOD project. They include physically and biogeochemically relevant parameters that can be directly compared with climate model simulations and for which reasonable spatiotemporal coverage could be achieved. All data have been compiled on a single depth scale per sediment core allowing for robust alignment of the various parameters and for reproducible updating and changing of chronologies. Whenever possible, the synthesis contains raw observational data together with the inferred data to enable revision of the latter. All data are associated with rich metadata to facilitate interpretation and for comprehensive uncertainty estimates. The synthesis is designed to facilitate evaluation of transient climate model simulations, but the data product can also be used to study climate change over the last glacial–interglacial cycle directly, independent from climate model simulations. Whilst the intended use of the synthesis has remained unchanged, version 1 was limited in spatial and temporal coverage, thus warranting an update to increase its usefulness. This article describes this update of the data synthesis and complements an earlier description of the data product (Jonkers et al., 2020).

2 What's new

Briefly, version 2 of the PALMOD 130k marine palaeoclimate data synthesis contains 1736 additional time series from 332 individual sites, markedly increasing both the spatial and temporal coverage of the data. Age-depth models presented in version 1 that were based on radiocarbon ages are revised using calendar ages calculated with the IntCal20 radiocarbon calibration curve (Hogg et al., 2020; Reimer et al., 2020). In addition, we harmonised seawater temperature estimates based on geochemical proxies and provide uncertainty estimates that are consistent and comparable across the different time series. This version also corrects several errors in the metadata associated with time series included in version 1 and contains two additional metadata fields to facilitate filtering of the data. Finally, we made slight changes in the structure of the data to prevent errors during use.

3 Methods

3.1 Data synthesis strategy

The synthesis strategy of this version follows that of the previous version (Jonkers et al., 2020). We compiled time series from marine sediment cores that contained benthic or planktonic foraminifera stable oxygen and carbon isotope data, estimates of seawater temperature and information on total organic carbon, opal and carbonate content. We followed a sequential updating strategy guided by the availability of time series with information needed to develop an age depth model for the sediment cores of interest. Version 1 contained time series dated with radiocarbon and benthic foraminifera δ¹⁸O. Version 2 integrates two steps in the update process, it contains time series with any of the parameters of interest with age control solely based on benthic foraminifera δ¹⁸O and it contains time series that had at least an estimate of seawater temperature and age control based on radiocarbon or other absolute dates. Data was compiled from public repositories following the two-step approach of bulk data download and subsequent filtering out of relevant time series as described in Jonkers et al. (2020). To increase the number of time series from the Indo-Pacific domain we made 19 previously unpublished datasets available on PANGAEA and included them here. Given the lack of metadata associated with most data in public repositories, metadata was in general manually compiled from the publications associated with the datasets.

3.2 Age-depth modelling

All time series in this compilation have updated and harmonised age-depth models with consistent estimates of uncertainty. The updated age-depth models are based on the most recent regional benthic foraminifera δ¹⁸O stacks and radiocarbon calibration curve, together with estimates of marine reservoir ages that are globally consistent and physically plausible. We follow the same hybrid strategy of relying on both absolute age control points (e.g. radiocarbon ages, tephra horizons) and hypothesis-based age modelling as in version 1.

For this version, benthic foraminifera δ¹⁸O were manually aligned to the relevant regional stacks of Lisiecki and Stern (2016) using mainly the marine isotope stage boundaries as tie points for consistency among the time series and to avoid inflating the chronological confidence. In addition, all radiocarbon ages have been recalibrated to calendar years using IntCal20 (Hogg et al., 2020; Reimer et al., 2020) and model-based estimates of the reservoir age (Butzin et al., 2017). The age-depth models were constructed using the Bayesian algorithm BACON (Blaauw and Christen, 2011) implemented in PDV (Langner and Mulitza, 2019) and are presented as separate elements in the compilation. To avoid extrapolation, we only provide age models for the core depths bracketed by age control points. Age-depth models presented in the original studies may thus extend further to either side. The age-depth information includes mean, median and the 95 % confidence interval of the ages as well as the necessary information to rerun the age-depth models in BACON and 1000 posterior age-depth models that can be used for further analysis.

3.3 Seawater temperature estimates

In order to homogenise seawater temperature estimates and to provide a consistent estimate of their uncertainty associated with the calibration and, in case of planktonic $\mathrm{Mg}/\mathrm{Ca}$ ratios, of the confounding effects of salinity and pH, temperature estimates were recalculated. For seawater temperatures based on the alkenone unsaturation index we used the BAYSPLINE calibration (Tierney and Tingley, 2018) and applied the suggested prior for the standard deviation of sea surface temperature of 5 °C. The resulting estimates represent in general annual mean sea surface temperature. However, for specific regions the temperatures are seasonal averages (Tierney and Tingley, 2018). This seasonality of the estimate is indicated in the metadata.

Whenever available, or whenever values could be calculated from inferred seawater temperatures, planktonic foraminifera $\mathrm{Mg}/\mathrm{Ca}$ ratios were converted to calcification temperature using MgCaRB (Gray and Evans, 2019). In this way we account for the effects of salinity and pH on the $\mathrm{Mg}/\mathrm{Ca}$ ratios based on sensitivities determined from culture studies. However, as the past salinity and pH variability is unknown for most time series in the synthesis, we, like Gray and Evans (2019) and Tierney et al. (2019), approximate the temporal evolution of salinity using a scaling to sea level and of pH based on atmospheric pCO₂ from ice cores and the constant disequilibrium from atmospheric CO₂ (ΔpCO₂) and alkalinity fixed at pre-industrial and modern levels respectively. For each site, we obtained modern sea surface salinity and temperature (needed for the calculation of ΔpCO₂) from the World Ocean Atlas 2023 (Reagan et al., 2023) from within a circle with a radius of 100 km. Carbonate system parameters (total dissolved inorganic carbon (DIC), its anthropogenic contribution (DIC_anthro) and total alkalinity were derived from the GLODAPv2 gridded data product (Lauvset et al., 2016) in the same way. Since GLODAPv2 does not cover the Gulf of Mexico and the Caribbean Sea, we used the median surface DIC and alkalinity from the GLODAPv2.2023 (Olsen et al., 2016) adjusted cruise data from the area between 10 and 30° N west of 80° W (Barbero et al., 2016a, b, 2019; Jiang et al., 2020; Salisbury, 2017; Salisbury and Shellito, 2019; Wanninkhof et al., 2007, 2014) for the affected time series. For simplicity we assumed global mean DIC_anthro for those sites. Local ocean pCO₂ was subsequently calculated using pre-industrial DIC (DIC-DIC_anthro), alkalinity, salinity and temperature using SeaCarb (Gattuso et al., 2024). Finally, ΔpCO₂ was calculated as the difference between the calculated oceanic pCO₂ and pre-industrial atmospheric pCO₂ (280 ppm). For the species Globigerinoides ruber, Trilobatus sacculifer, Globigerina bulloides we used the sensitivity of $\mathrm{Mg}/\mathrm{Ca}$ to temperature, salinity and pH as provided in Gray and Evans (2019). For Neogloboquadrina pachyderma and N. incompta we used the temperature and salinity sensitivity as used in Boscolo-Galazzo et al. (2025) and ignored the effect of pH as this is poorly constrained for these species. We used the generic calibration for all other species. To account for Mg loss in foraminifera tests that were treated by reductive and oxidative cleaning, we multiplied the respective $\mathrm{Mg}/\mathrm{Ca}$ ratios by 1.1 to make them comparable with ratios measured on samples that only underwent oxidative cleaning (Barker et al., 2003).

For seawater temperatures derived from both proxies, estimates are presented as the median and 95 % confidence intervals. The recalculated values are indicated using the strings “recalc_UKSST” or “recalc_mgcaSST” and accompanied by a note containing the text “Recalculated for PALMOD”. For further analysis, 1000 ensemble time series of seawater temperature are provided separately.

Unlike for the two proxy types mentioned above, seawater temperature estimates based on microfossil assemblages were not harmonised. This is because in too many cases the calibration that was used to generate the reported temperature estimates and/or the fossil abundance data on which the estimates are based are not publicly available.

4 Structure, terms and format

The structure of the synthesis has slightly changed since version 1. The main change is the inclusion of the revised age information directly with the proxy data. These were kept separate in version 1 to highlight the fact that observations on the (core) depth scale are static, whereas the age-depth model can be updated or revised as per user demand. However, this separation into raw and derived data rendered use of the data more complex and we have hence decided to include the revised age data directly with the proxy data. The second change in the structure is the addition of ensembles of seawater temperature time series.

The synthesis is organised by site, which usually is the sediment core from which the data were extracted, but it may also refer to a spliced record comprising multiple sediment cores. Depending on the availability of alkenone or $\mathrm{Mg}/\mathrm{Ca}$-derived seawater temperature estimates, each site is associated with information on seven or eight themes. Names in square brackets below are those used in the data files.

• Geographic data [site] includes the position of the site in decimal degrees and its water depth in m and, new in this version, the ocean basin in which the site is located. The basin assignment is based on the Global Oceans and Seas dataset, version 1 (Flanders Marine Institute (VLIZ), Belgium, 2021). This theme also contains additional notes that apply to the entire site.

• Metadata [meta] contains the original and standardised variable name, units where applicable, information about the origin of the data and where applicable information about the proxy sensor and the seasonal or vertical attribution (as specified in the original publications) and the calibration. A complete list of metadata terms is provided in Table 1. Metadata are only included if they were available in the dataset or could be obtained from the accompanying publication. Not all fields are relevant for all parameters (e.g. species is not defined for parameters that are not based on measurements on organisms) and fields that are completely empty for a site are excluded. The metadata now also includes a unique identifier for each time series to facilitate cross-referencing.

• Chronological data [chron] includes the information to revise the age-depth model of the site. It contains raw radiocarbon ages, ages based on tephra layers and palaeomagnetics, as provided by the data generators and the ages of the tie points identified for this study. The calendar ages of all tie points are based on the approach outlined above (section Age-depth modelling). Information about the source and the citation of the chronological information is provided. A full list of chronological data terms is provided in Table 2.

• Proxy data [data] are presented on a common depth scale. Original ages and original seawater temperature estimates are included, updated seawater temperature estimates are flagged in the notes field using the string “Recalculated for PALMOD”. The standardised parameter names are provided in Table 3.

• Age-depth model ensembles [AgeEnsemblesBACON]. 1000 posterior age-depth models inferred from sediment accumulation rates as modelled using BACON.

• Seawater temperature ensembles [SurfaceTempEnsembles]. 1000 sea water temperature time series for each $U_{\mathrm{37}}^{K^{\prime }}$ or $\mathrm{Mg}/\mathrm{Ca}$ time series. The ensembles are accompanied by metadata describing the algorithm settings and information about the confidence of the resulting temperature estimates (expressed as convergence, see Tierney and Tingley (2018) for further explanation) and details on the species in the case of $\mathrm{Mg}/\mathrm{Ca}$.

Table 1 Metadata fields.

^* Original parameter names are preserved even when they contain errors. These errors have been corrected in the standardised parameter name and additional metadata.

Each site also contains information about the version, import date and date at which the age-depth model was last updated. These fields are largely for internal use.

The data are available in serialised R format (^*RDS) and structured as a list where the elements contain the information about the themes indicated above. RDS files can be easily read using the open source software R. To ensure interoperability the data product is also available as (simple) JSON formatted data following the LiPD structure and vocabulary. Resources for reading and querying the data are provided in the Sect. 7 Usage notes.

Table 2 Chronology data fields.

Table 3 Standardised parameter names.

5 Description/contents

5.1 Spatial data coverage

Version 2 of the PALMOD marine palaeoclimate data synthesis contains 2286 time series with one of the eight synthesised climate-relevant variables each, originating from 475 unique sites (Fig. 1, Table 4). The majority of the newly added time series in this version stem from the tropics. Most time series are of planktonic foraminifera δ¹⁸O (n= 482), with often multiple time series from different species (each with potentially different vertical and seasonal recording preferences) per site. The number of time series with benthic foraminifera δ¹⁸O and with near sea surface temperature closely follow the number of time series with benthic δ¹⁸O (n= 468 and n= 423, respectively). Time series of the other variables are considerably fewer in number. For all variables but biogenic silica content, the highest density of sites is in the Atlantic Ocean. Coverage in the Pacific and the southern hemisphere remains relatively poor, reflecting both historical research focus and the (logistical) challenges of obtaining sediment cores far offshore. In general, sites tend to be located close to the continents, where higher sedimentation rates facilitate the recovery of records with higher temporal resolution.

Figure 1 Spatial distribution of sites with time series of the eight parameters. Locations are coloured by the version in which they were first included (version 2 also includes all data of version 1). The total number of sites and time series are indicated in brackets.

5.2 Temporal data coverage

Compared to version 1, the temporal coverage has approximately doubled over the 130 kyr time frame and for all variables (Fig. 2). The temporal distribution of data reflects research focus on the last deglaciation, with the highest data density between 10–25 ka BP. Importantly, the coverage does not markedly drop off beyond approximately 40 ka BP, facilitating continuous analysis across the entire last glacial cycle. This coverage pattern is consistent across all variables and has not substantially changed in version 2.

Figure 2 Temporal distribution of temporal coverage of each parameter. Lines show the number of time series per 5 kyr bin, colour show the total number of time series for each version. Like in the previous version, version 2 shows highest coverage during the last deglaciation. Note the different scaling on the y axes.

The length of the time series shows a clear bimodal distribution across all variables (Fig. 3). Most time series are shorter than 40 kyr or longer than 110 kyr, reflecting the research focus on investigating either the transition from the last glacial to the Holocene, or glacial-interglacial cycles on long(er) time scales.

Figure 3 Distribution of time series length split by parameter. Time series between 40 and 110 kyr long are uncommon across all parameters.

The temporal resolution of the time series varies considerably, but the median temporal resolution of the time series tends to be below 3 kyr (Fig. 4). In general, the highest temporal resolution of the time series is in the youngest part of the 130 kyr time frame (Fig. 4). Even though the resolution decreases with age, this decrease generally stops or reduces around 75 ka BP. The resolution of the time series is, to a first order, related to their length (Fig. 5), reflecting both the challenge of obtaining long sediment cores in high accumulation settings and the amount of work involved in generating the data.

Figure 4 Temporal resolution (observations per kyr) of the time series split by parameter. The general trend shows decreasing resolution until approximately 75 ka BP, after which resolution remains more or less constant. Even though there is considerable spread, the median resolution is below 3 kyr across all parameters. Box-whisker plots show distribution of median resolution per 5 kyr bin; thick lines indicate median values, boxes show interquartile range and whiskers extend to 1.5 times this range.

Figure 5 Longer time series tend to have lower resolution. Cross plot of median resolution (time between consecutive observations) and time series length. The linear fit (black line) is shown to highlight the general trend; 14 time series shorter than 1 kyr are not shown.

5.3 Chronology

Nearly 90 % of the time series have an age-depth model based on exclusively radiocarbon dating and tuning of benthic foraminifera δ¹⁸O (Fig. 6a). Only five time series have age-depth models entirely based on other – e.g. tephra, biostratigraphy or palaeomagnetics – age controls. The age uncertainty shows to some degree the nature of the dating methods, it tends to be low up to approximately 20 ka BP, reflecting the prevalence of radiocarbon dating within this interval (Fig. 6b). Beyond ca. 20 ka BP, the age uncertainty shows the structure of the benthic foraminifera δ¹⁸O tuning targets with reduced uncertainties near the marine isotope stage boundaries that were used as tie points.

Figure 6 Type of age control tie points across all sites (a), age uncertainty across the time series in the dataproduct (b) and differences between the original and updated age-depth models (c). The category “other” in panel a includes tephra, palaeomagnetics and biostratigraphy. Low age uncertainties (b) up to approximately 25 ka BP reflect the prevalence of radiocarbon ages within this interval. The age uncertainty in the older parts of the time series tends to be lower near the boundaries of the marine isotope stages that were used in the construction of the age depth models. Box-whisker plots show the distribution of median age uncertainty (95 % confidence interval) across the sites per 5 kyr bin. Further explanation of the plots is provided in Fig. 4. Colours in c show whether or not the difference between the age estimates is within the 95 % CI of the revised age-depth model. In 2 % of the data points the difference is outside the range shown here.

The revised age-depth models differ in many cases from the original age-depth models (Fig. 6c). However, these differences are small, the median difference (revised – original) is −0.07 kyr (−3.28–5.24, 95 % CI) and for 74 % of all data points the difference is within the uncertainty of the revised age-depth model. These differences are to some degree due to the different approaches used to model the age-depth relationship and/or to differences in the radiocarbon calibration curve, the applied reservoir age, the tuning parameters and the tuning targets.

5.4 Near sea surface temperature

The majority of the near sea surface temperature time series (n= 173) is based on planktonic foraminifera $\mathrm{Mg}/\mathrm{Ca}$ (Fig. 7a). Near sea surface temperature estimates based on alkenone unsaturation indices and microfossil assemblages (predominantly planktonic foraminifera) make up approximately equal proportions (n= 119 and 115, respectively). Only 16 time series are based on other proxies, including TEX₈₆ (Schouten et al., 2002) and the long chain diol index (Rampen et al., 2012). Twenty-nine sites have near sea surface temperature estimates based on the $\mathrm{Mg}/\mathrm{Ca}$ of more than one species of planktonic foraminifera and 55 sites have near sea surface temperature time series based on more than one proxy (Fig. 7a), thus potentially allowing for comparison and/or the reconstruction of seasonal or vertical temperature gradients. The uncertainty of the revised near sea surface temperature estimates (1 SD) remains approximately constant through time and its median value across all samples is 0.93 °C (0.64–2.44 °C, 95 %).

Figure 7 Number of the near sea surface temperature time series sorted by proxy (a) and a Venn diagram showing the overlap among records based on different proxies across the sites (b; numbers indicated). In total 55 sites have near sea surface temperature data based on more than one proxy.

6 Usage notes

General instructions on data usage and potential applications can be found in Jonkers et al. (2020). The PALMOD synthesis contains data about sediment composition (calcium carbonate, organic carbon and biogenic silica content) that is relatively straightforward to interpret. The data product also contains information derived from measurements on biological material (foraminifera stable carbon and oxygen isotope ratios) and estimates of seawater derived indirectly from the chemical or species composition of microfossils or from biomarkers. The biological origin of the data renders its interpretation less straightforward as environmental preferences of the organisms may affect the recorded climate signal. A detailed description of the proxy sensors and the intricacies of the proxies is provided in the data descriptor of version 1 (Jonkers et al., 2020). We encourage users to read the relevant section. A general proxy system modelling framework for marine sediment time series can e.g. be found in Dolman and Laepple (2018). In this section we focus on the main advance compared to version 1 that is relevant to usage: the harmonisation of seawater temperature estimates.

Estimates of near sea surface temperature based on planktonic foraminifera $\mathrm{Mg}/\mathrm{Ca}$ and $U_{\mathrm{37}}^{K^{\prime }}$ have been updated and harmonised whenever possible. For $U_{\mathrm{37}}^{K^{\prime }}$, these updated estimates required relatively little user input, but the use of different priors may affect the results. Since planktonic foraminifera $\mathrm{Mg}/\mathrm{Ca}$ is not only affected by seawater temperature, the inference of $\mathrm{Mg}/\mathrm{Ca}$-based temperatures is more complex. Estimates of salinity and pH are indirect and, by necessity, rather crude as they ignore existing, but unknown, spatiotemporal variation. Importantly, the estimates of salinity and pH depend on the age and updates to the age-depth model therefore require recalculation of the inferred temperatures. The influence of these nuisance factors can in principle also be corrected using measured estimates of ocean pH based on boron isotope ratios (Gray and Evans, 2019) or indirectly by using oxygen isotopes (Morley et al., 2024), but this requires additional data that is not available for all time series. We have hence chosen a correction approach that is consistent across all time series. Whilst for the species Globigerinoides ruber, Trilobatus sacculifer, Globigerina bulloides, Neogloboquadrina pachyderma and N. incompta the sensitivity of $\mathrm{Mg}/\mathrm{Ca}$ to temperature, salinity and pH is well constrained, this is not the case for other, generally deeper dwelling species. For these other species we have used the generic calibration. Care thus needs to be taken with the temperature estimates based on this calibration and we encourage the use of anomalies, rather than reliance on absolute values. To allow users to compare the harmonised seawater temperature estimates with the values provided by the data generators, the original temperature values are included in the data product. With regard to the estimated uncertainty of the inferred temperatures, it is important to note that the range of possible nSST trajectories may be overestimated as temporal autocorrelation in the temperature and its uncertainty are ignored in the calibration. Whether such autocorrelation would be preserved in sedimentary time series and how large the effect would be, remains to be investigated. For this, and other reasons, users may also want to use different calibration schemes. To this end, the synthesis also contains the raw $U_{\mathrm{37}}^{K^{\prime }}$ and $\mathrm{Mg}/\mathrm{Ca}$ values.

Seawater temperatures based on microfossil assemblage composition are included in the synthesis only in the form reported by the authors. Because of the fragmentary nature of the required metadata, the originally reported estimates could not be updated to a common standard and are associated with uncertainty estimates only where the data generators provided them. From the accompanying publications it is not always clear how the uncertainty has been derived and the methods certainly vary among the different studies because of personal preferences and methodological progress on how to deal with issues like spatial autocorrelation. In light of new insights regarding the uncertainty of seawater temperature estimates based on microfossil assemblages (Telford et al., 2013; Trachsel and Telford, 2016), the reported uncertainties are likely too optimistic. Interested users can refer to Jonkers et al. (2023) for an example of comprehensive estimates of calibration uncertainty associated with planktonic foraminifera assemblage derived seawater temperatures.

Whenever publicly available, the original age-depth models are included in the data product, but they almost universally lack uncertainty estimates. Still, these original age estimates can be used for a comparison of the revised and original age models. Future updates to the radiocarbon calibration curve, to the reservoir age estimates, or the stacks may require updating the age-depth models. The raw chronology data provided in the data product enables this updating in an automated fashion. Please note that the $\mathrm{Mg}/\mathrm{Ca}$-based temperature estimates depend on the age of the samples for a correction of the effects of salinity and pH. The temperature estimates thus need to be updated if the age-depth model is revised. Uncertainty estimates on the chronology are to some degree method dependent, but the Bayesian approach we utilised appears to provide reasonable results (Trachsel and Telford, 2017). However, our approach ignores additional uncertainty caused by bioturbation and uncertainty related to alignment of the benthic foraminifera δ¹⁸O records. Recent work suggests ways to incorporate these sources of uncertainty (Lee et al., 2023). Other approaches may thus provide different results and the user may wish to compare age-depth models using different age-depth modelling strategies like for instance done by Comas-Bru et al. (2020).

The availability of age and nSST ensembles offers the possibility to estimate the joint chronological and calibration uncertainty of the seawater temperature estimates. A possible way to do so is through combining the age and nSST ensembles in a Markov Chain Monte Carlo approach (Cartapanis et al., 2022; Khider et al., 2017), by randomly joining ensembles (Fig. 8a). However, the large number of ensembles potentially renders this approach excessive and computationally expensive as 10⁶ unique combinations of the ensembles are possible. A preliminary analysis suggests that 500–1000 random combinations of age and nSST ensembles are sufficient to capture the combined age and calibration uncertainty (Fig. 8b). We caution however against universal application of this number as the distributions of age and calibration uncertainty likely vary by time series and different research questions may require different approaches. We hence encourage users to carefully consider the chosen number of ensemble combinations. Alternatively, the joint uncertainty may be quantified analytically, but such an approach needs to consider the potential complexity in the probability density distributions of chronological and calibration uncertainty.

Figure 8 An example of joint estimation of chronological and calibration uncertainty of nSST timeseries. Panel (a) shows the original nSST time series of site A7 in the North Pacific that is based on the $\mathrm{Mg}/\mathrm{Ca}$ ratio of Globigerinoides ruber albus (Sun et al., 2005) as well as the updated time series with uncertainty estimated as described in the main text. The solid line shows the median nSST and the light and dark grey are 90 % and 50 % central percentiles, respectively. Panel (b) shows the median uncertainty of all nSST (90 %) values of the entire time series as a function of the number of ensemble pairs (age and nSST). This example shows that out of 1 000 000 possible pairs about 500–1000 are sufficient to characterise the uncertainty of this time series. The number of ensemble pairs required for other time series, or other purposes, needs to be evaluated on a case-by-case basis.

Some example scripts for filtering the data product by location, length, resolution and age control are available from the lead author's GitHub repository (https://github.com/lukasjonkers/PALMODutils, last access: 22 January 2026). Example code to estimate the uncertainty of the nSST time series based on the age and temperature ensembles is available at the same location. An extensive documentation of the LiPD format is available at https://lipdverse.org (last access: 23 May 2025) and utilities to interact with LiPD files in R are available at https://github.com/nickmckay/lipdR (last access: 20 January 2026) and a user guide to the lipdR package is available at https://nickmckay.org/lipdR/ (last access 20 January 2026). Python users can use the PyLiPD package (https://pylipd.readthedocs.io/en/latest/, last access: 20 January 2026), which is documented at https://linked.earth/pylipdTutorials/intro.html (last access: 20 January 2026). Resources for Matlab users are available at https://github.com/nickmckay/LiPD-utilities (last access: 23 May 2025).

Whilst we have done our best to address errors and omissions and aimed for standardisation across the entire data product, it comes with the guarantee of remaining errors and inconsistencies. Users are encouraged to report those to the lead author by email so they can be corrected. Updates to the data synthesis will be published on PANGAEA and linked to the current version. Minor updates with corrections only will not be associated with a new data descriptor. The updating and versioning strategy is described in detail in Jonkers et al. (2020).

7 Data availability

The PALMOD 130k marine palaeoclimate data synthesis version 2 is freely available at https://doi.org/10.1594/PANGAEA.984602 (Jonkers et al., 2026). The data can also be visualised and downloaded in LiPD format from https://lipdverse.org/PalMod/current_version/ (last access: 30 September 2025). Python and Matlab serialisations of the data as well as csv files of the palaeoclimate and chronological data can also be downloaded from the lipdverse website. We explicitly encourage users to also cite the original data when using this data product to provide credit to the data generators. For instance, when using a selection of the time series from this data product, we recommend including a statement like “We selected time series from the PALMOD 130k marine palaeoclimate synthesis version 2 (Jonkers et al., 2026). The time series that meet our selection criteria, including their original publications, are detailed in the table below”. See Weitzel et al. (2024) for an additional example and Smith et al. (2023) for a general discussion on the issue of crediting data generators.

8 Recommendations for palaeoclimate data stewardship

This synthesis would not have been possible without good data stewardship of the marine palaeoclimate science community. For many decades researchers have been sharing their research data through (dedicated) publicly accessible data repositories. Even so, this synthesis makes a lot of invaluable metadata and chronological information publicly available for the first time. This is because such metadata and chronology data are often not directly associated with the primary proxy data, but reported in accompanying publications instead. When these publications are publicly accessible, this important information to interpret and reuse the data can be retrieved with considerable additional effort. However, more often than not, the publications are not publicly available and the information remains hidden behind the paywalls of publishers, rendering the data difficult to reuse.

Community guidelines on how to report general palaeoclimate data are available (Khider et al., 2019), as are more specific guidelines for microfossil assemblage data (Jonkers et al., 2025). However, legacy datasets invariably do not adhere to such guidelines, thus necessitating (community) effort to increase the reusability of these data. For new datasets, we recommend data sharing not only from the point of view of reproducing single studies, but explicitly consider reuse when sharing data and ensure that that research data files are scientific works that are comprehensible in themselves. In this way, reuse and reinterpretation of scientific data and knowledge is greatly facilitated, allowing our field to address new questions that cannot be answered using a single palaeoclimate time series.

With the National Climate Data Center (NCDC) at NOAA and PANGAEA the palaeoclimate community has access to at least two dedicated data repositories that enforce standardisation and perform data curation. This is a tremendously advantageous position to facilitate the reusability of palaeoclimate data and we explicitly encourage data generators to continue sharing their data through these recognised and specific data repositories. Nevertheless and despite the existence of the LinkedEarth ontology (https://linked.earth/ontology/, last access: 20 January 2026) and the PaST thesaurus (Morrill et al., 2021), the lack of adaptation of a single standardised vocabulary and consistent ontology across the repositories remains a hurdle to reusability. The PaST thesaurus is used at NCDC, but standardisation is only performed under the hood at PANGAEA (Felden et al., 2023), where structured and standardised data can be accessed in xml format. Synthesis efforts benefit from such standardisation, but increased harmonisation across the repositories, or tools to do so, would further facilitate synthesis efforts and the reusability of data.

The value of this data product arguably not only lies in the standardisation of the format and vocabulary, but also in the updating and standardisation of the age-depth models and of the seawater temperature estimates. Such updating and standardisation is only possible when raw chronology and proxy data are available on a depth scale. Radiocarbon data are often publicly available or can, with additional effort, be scraped from publications. However, information on aligned chronologies, including ages and uncertainty on tuning tie points is seldom reported, rendering reevaluation of such chronologies difficult. Regarding proxy data, measured $\mathrm{Mg}/\mathrm{Ca}$ and $U_{\mathrm{37}}^{K^{\prime }}$ can, provided the calibration equation is known, be calculated from reported seawater temperatures, but ambiguities are likely to remain and the chance of deriving erroneous values is non-negligible (especially when the calibration equation is not explicitly included in the data file, or provided in the associated publication). For seawater temperatures based on microfossil assemblage the situation is different. Raw assemblage data are not as often shared as geochemical proxy data and due to the multi-dimensional character of the data, raw assemblage data cannot be inferred from the temperature estimates. This not only prevents revising temperature estimates reported without the underlying raw count data, but also prevents reuse of such assemblage data in other contexts, such as for estimates of past biodiversity change (e.g. Strack et al., 2024). This is unfortunate given the tremendous effort that went into generating such species assemblage data. In our opinion, the amount of work and the relevance of such data is so large that it would justify a coordinated international action to rescue the unpublished microfossil counts underlying seawater temperature reconstructions.

Table 4 Sites included in the PALMOD 130k marine palaeoclimate data synthesis version 2.