The global atmospheric GCM ECHAM5 (Roeckner et al., 2003) has been developed at the Max Planck Institute for Meteorology and is based on the spectral weather forecast model of the ECMWF (Simmons et al., 1989). In the context of palaeoclimate applications, the model has been used mostly at lower resolution (T31, ca. 3.75∘ × 3.75∘; T63, ca. 1.9∘ × 1.9∘ in the case of Feng et al., 2016, and T106 in the case of Li et al., 2017 and Feng and Poulsen, 2016). The studies performed are not limited to the last millennium (e.g. Jungclaus et al., 2010) but also include research in the field of both warmer and colder climates, at orbital (e.g. Gong et al., 2013; Lohmann et al., 2013; Pfeiffer and Lohmann, 2016; X. Zhang et al., 2013, 2014; Wei and Lohmann, 2012) and tectonic timescales (e.g. Knorr et al., 2011; Stepanek and Lohmann, 2012), and under anthropogenic influence (Gierz et al., 2015).

Here, the ECHAM5 simulations were conducted at a T159 spatial resolution (horizontal grid size ca. 80 km × 80 km at the Equator) with 31 vertical levels (between the surface and 10 hPa). This high model resolution is admittedly not required for all of the climatological questions investigated in this study, and it should be noted that the skill of GCMs in predicting orographic precipitation remains limited at this scale (e.g. Meehl et al., 2007). However, simulations were conducted at this resolution so that future work can apply the results in combination with different dynamical and statistical downscaling methods to quantify changes at large catchment to orogen scales. The output frequency is relatively high (1 day) to enhance the usefulness of our simulations as input for landscape evolution and other models that may benefit from daily input. The simulations were conducted for five different time periods: present-day (PD), PI, MH, LGM, and PLIO.

A PD simulation (not shown here) was used to establish confidence in the model performance before conducting palaeosimulations and has been compared with the following observation-based datasets: European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses (ERA40, Uppala et al., 2005), National Centers for Environmental Prediction and National Center for Atmospheric Research (NCEP/NCAR) reanalyses (Kalnay et al., 1996; Kistler et al., 2001), NCEP Regional Reanalysis (NARR; Mesinger et al., 2006), the Climate Research Unit (CRU) TS3.21 dataset (Harris et al., 2013), High Asia Refined Analysis (HAR30; Maussion et al., 2014), and the University of Delaware dataset (UDEL v3.01; Legates and Wilmott, 1990). (See Mutz et al., 2016, for a detailed comparison with a lower-resolution model).

The PI climate simulation is an ECHAM5 experiment with PI (reference year 1850) boundary conditions. Sea surface temperatures (SSTs) and sea ice concentration (SIC) are derived from transient coupled ocean–atmosphere simulations (Lorenz and Lohmann, 2004; Dietrich et al., 2013). Following Dietrich et al. (2013), greenhouse gas (GHG) concentrations (CO2 : 280 ppm) are taken from ice-core-based reconstructions of CO2 (Etheridge et al., 1996), CH4 (Etheridge et al., 1998) and N2O (Sowers et al., 2003). Sea surface boundary conditions for the MH originate from a transient, low-resolution, coupled atmosphere–ocean simulation of the MH (6 ka) (Wei and Lohmann, 2012; Lohmann et al., 2013), where the GHG concentrations (CO2 : 280 ppm) are taken from ice core reconstructions of GHGs by Etheridge et al. (1996, 1998) and Sowers et al. (2003). GHG concentrations for the LGM (CO2 : 185 ppm) have been prescribed following Otto-Bliesner et al. (2006). Orbital parameters for the MH and LGM are set according to Dietrich et al. (2013) and Otto-Bliesner et al. (2006), respectively. LGM land–sea distribution and ice sheet extent and thickness are set based on the PMIP III (Palaeoclimate Modelling Intercomparison Project, phase 3) guidelines (elaborated on by Abe-Ouchi et al., 2015). Following Schäfer-Neth and Paul (2003), SST and SIC for the LGM are based on GLAMAP (Sarnthein et al., 2003) and CLIMAP (CLIMAP project members, 1981) reconstructions for the Atlantic Ocean and Pacific and Indian oceans, respectively. Global MH and LGM vegetation is based on maps of plant functional types by the BIOME 6000 Palaeovegetation Mapping Project (Prentice et al., 2000; Harrison et al., 2001; Bigelow et al., 2003; Pickett et al., 2004) and model predictions by Arnold et al. (2009). Boundary conditions for the PLIO simulation, including GHG concentrations (CO2 : 405), orbital parameters and surface conditions (SST, SIC, sea land mask, topography, and ice cover) are taken from the PRISM (Pliocene Research, Interpretation and Synoptic Mapping) project (Haywood et al., 2010; Sohl et al., 2009; Dowsett et al., 2010), specifically PRISM3D. The PLIO vegetation boundary condition was created by converting the PRISM vegetation reconstruction to the JSBACH plant functional types as described by Stepanek and Lohmann (2012), but the built-in land surface scheme was used.

SST reconstructions can be used as an interface between oceans and atmosphere (e.g. Li et al., 2017) instead of conducting the computationally more expensive fully coupled atmosphere–ocean GCM experiments. While the use of SST climatologies comes at the cost of capturing decadal-scale variability, and the results are ultimately biased towards the SST reconstructions the model is forced with; the simulated climate more quickly reaches an equilibrium state and the means of atmospheric variables used in this study do no change significantly after the relatively short spin-up period. The palaeoclimate simulations (PI, MH, LGM, PLIO) using ECHAM5 are therefore carried out for 17 model years, of which the first 2 years are used for model spinup. The monthly long-term averages (multi-year means for individual months) for precipitation, temperature, and precipitation and temperature amplitude, i.e. the mean difference between the hottest and coldest months, have been calculated from the following 15 model years for the analysis presented below.

For further comparison between the simulations, the investigated regions were subdivided (Fig. 1). Western South America was subdivided into four regions: parts of tropical South America (80–60∘ W, 23.5–5∘ S); temperate South America (80–60∘ W, 50–23.5∘ S); tropical Andes (80–60∘ W, 23.5–5∘ S; high-pass filtered), i.e. most of the Peruvian Andes, Bolivian Andes, and northernmost Chilean Andes; and temperate Andes (80–60∘ W, 50–23.5∘ S, high-pass filtered). South Asia was subdivided into three regions: tropical South Asia (40–120∘ E, 0–23.5∘ N), temperate South Asia (40–120∘ E, 23.5–60∘ N), and high-altitude South Asia (40–120∘ E, 0–60∘ N; high-pass filtered).

Our approach of using a single GCM (ECHAM5) for our analysis is motivated by, and differs from, previous studies where inter-model variability exists from the use of different GCMs due to different parameterisations in each model. The variability in previous inter-model GCM comparisons exists despite the use of the same forcings (e.g. see results highlighted in IPCC AR5). Similarities identified between these palaeoclimate simulations conducted with different GCMs using similar boundary conditions can establish confidence in the models when in agreement with proxy reconstructions. However, differences identified in inter-model GCM comparisons highlight biases by all or specific GCMs, or reveal sensitivities to one changed parameter, such as model resolution. Given these limitations of GCM modelling, we present in this study a comparison of a suite of ECHAM5 simulations to proxy-based reconstructions (where possible) and, to a lesser degree, comment on general agreement or disagreement of our ECHAM5 results with other modelling studies. A detailed inter-model comparison of our results with other GCMs is beyond the scope of this study and better suited for a different study in a journal with a different focus and audience. Rather, by using the same GCM and identical resolution for the time slice experiments, we reduce the number of parameters (or model parameterisations) varying between simulations and thereby remove potential sources of error or uncertainty that would otherwise have to be considered when comparing output from different models with different parameterisations of processes, model resolution, and in some cases model forcings (boundary conditions). Nevertheless, the reader is advised to use these model results with the GCM's shortcoming and uncertainties in boundary condition reconstructions in mind. For example, precipitation results may require dynamical or statistical downscaling to increase accuracy where higher-resolution precipitation fields are required. Furthermore, readers are advised to familiarise themselves with the palaeogeography reconstruction initiatives and associated uncertainties. For example, while Pliocene ice sheet volume can be estimated, big uncertainties pertaining to their locations remain (Haywood et al., 2010).

The aim of the clustering approach is to group climate model surface grid boxes together based on similarities in climate. Cluster analyses are statistical tools that allow elements (i) to be grouped by similarities in the elements' attributes. In this study, those elements are spatial units, the elements' attributes are values from different climatic variables, and the measure of similarity is given by a statistical distance. The four basic variables used as climatic attributes of these spatial elements are near-surface (2 m) air temperature, seasonal 2 m air temperature amplitude, precipitation rate, and seasonal precipitation rate amplitude. Since monsoonal winds are a dominant feature of the climate in the South Asia region, near-surface (10 m) speeds of u wind and v wind (zonal and meridional wind components, respectively) during the monsoon season (July) and outside the monsoon season (January) are included as additional variables in our analysis of that region. Similarly, u-wind and v-wind speeds during (January) and outside (July) the monsoon season in South America are added to the list of considered variables to take into account the South American Monsoon System (SASM) in the cluster analysis for this region. The long-term monthly means of those variables are used in a hierarchical clustering method, followed by a non-hierarchical k-means correction with randomised regroupment (Mutz et al., 2016; Wilks, 2011; Paeth, 2004; Bahrenberg et al., 1992).

The hierarchical part of the clustering procedure starts with as many clusters as there are elements (ni), then iteratively combines the most similar clusters to form a new cluster using centroids for the linkage procedure for clusters containing multiple elements. The procedure is continued until the desired number of clusters (k) is reached. One disadvantage of a pure hierarchical approach is that elements cannot be recategorised once they are assigned to a cluster, even though the addition of new elements to existing clusters changes the clusters' defining attributes and could warrant a recategorisation of elements. We address this problem by implementation of a (non-hierarchical) k-means clustering correction (e.g. Paeth, 2004). Elements are recategorised based on the multivariate centroids determined by the hierarchical cluster analysis in order to minimise the sum of deviations from the cluster centroids. The Mahalanobis distance (e.g. Wilks, 2011) is used as a measure of similarity or distance between the cluster centroids since it is a statistical distance and thus not sensitive to different variable units. The Mahalanobis distance also accounts for possible multi-collinearity between variables.

The end results of the cluster analyses are subdivisions of the climate in the investigated regions into k subdomains or clusters based on multiple climate variables. The region-specific k has to be prescribed before the analyses. A large k may result in redundant additional clusters describing very similar climates, thereby defeating the purpose of the analysis to identify and describe the dominant, distinctly different climates in the region and their geographical coverage. Since it is not possible to know a priori the ideal number of clusters, k was varied between 3 and 10 for each region and the results presented below identify the optimal number of visibly distinctly different clusters from the analysis. Optimal k was determined by assessing the distinctiveness and similarities between the climate clusters in the systematic process of increasing k from 3 to 10. Once an increase in k no longer resulted in the addition of another cluster that was climatologically distinctly different from the others, and instead resulted in at least two similar clusters, k of the previous iteration was chosen as the optimal k for the region.

The cluster analysis ultimately results in a description of the geographical extent of a climate (cluster) characterised by a certain combination of mean values for each of the variables associated with the climate. For example, climate cluster 1 may be the most tropical climate in a region and thus be characterised by high precipitation values, high temperature values, and low seasonal temperature amplitude. Each of the results (consisting of the geographical extent of climates and mean vectors describing the climate) can be viewed as an optimal classification for the specific region and time. It serves primarily as a means for providing an overview of the climate in each of the regions at different times, reduces dimensionality of the raw simulation output, and identifies regions of climatic homogeneity that are difficult to notice by viewing simple maps of each climate variable. Its synoptic purpose is similar to that of the widely known Köppen–Geiger classification scheme (Peel et al., 2007), but we allow for optimal classification rather than prescribe classes, and our selection of variables is more restricted and made in accordance with the focus of this study.

This section describes the differences between simulated MH, LGM, and PLIO annual mean temperature anomalies with respect to PI shown in Fig. 2b, and PI temperature absolute values shown in Fig. 2a. Most temperature differences between the PI and MH climate are within -1 to 1 ∘C. Exceptions to this are the Hudson Bay, Weddell Sea, and Ross Sea regions, which experience warming of 1–3, 1–5, and 1–9 ∘C, respectively. Continental warming is mostly restricted to low-altitude South America, Finland, western Russia, the Arabian peninsula (1–3 ∘C), and subtropical North Africa (1–5 ∘C). Simulation results show that LGM and PLIO annual mean temperature deviate from the PI means the most. The global PLIO warming and LGM cooling trends are mostly uniform in direction, but the magnitude varies regionally. The strongest LGM cooling is concentrated in regions where the greatest change in ice extent occurs (as indicated in Fig. 2), i.e. Canada, Greenland, the North Atlantic, northern Europe, and Antarctica. Central Alaska shows no temperature changes, whereas coastal southern Alaska experiences cooling of ≤ 9 ∘C. Cooling in the US Pacific Northwest is uniform and between 11 and 13 ∘C. Most of high-altitude South America experiences mild cooling of 1–3 ∘C, 3–5 ∘C in the central Andes, and ≤ 9 ∘C in the south. Along the Himalayan orogen, LGM temperature values are 5–7 ∘C below PI values. Much of central Asia and the Tibetan Plateau cools by 3–5 ∘C, and most of India, low-altitude China, and South East Asia cools by 1–3 ∘C.

In the PLIO climate, parts of Antarctica, Greenland, and the Greenland Sea experience the greatest temperature increase (≤ 19 ∘C). Most of southern Alaska warms by 1–5 and ≤ 9 ∘C near McCarthy, Alaska. The US Pacific Northwest warms by 1–5 ∘C. The strongest warming in South America is concentrated at the Pacific west coast and the Andes (1–9 ∘C), specifically between Lima and Chiclayo, and along the Chilean–Argentinian Andes south of Bolivia (≤ 9 ∘C). Parts of low-altitude South America to the immediate east of the Andes experience cooling of 1–5 ∘C. The Himalayan orogen warms by 3–9 ∘C, whereas Myanmar, Bangladesh, Nepal, northern India, and northeastern Pakistan cool by 1–9 ∘C.

Notable differences occur between simulated MH, LGM, and PLIO annual mean precipitation anomalies with respect to PI shown in Fig. 3b and the PI precipitation absolute values shown in Fig. 3a. Of these, MH precipitation deviates the least from PI values. The differences between MH and PI precipitation on land appear to be largest in northern tropical Africa (increase ≤ 1200 mm a-1), along the Himalayan orogen (increase ≤ 2000 mm a-1), and in central Indian states (decrease) ≤ 500 mm. The biggest differences in western South America are precipitation increases in central Chile between Santiago and Puerto Montt. The LGM climate shows the largest deviation in annual precipitation from the PI climate, and precipitation on land mostly decreases. Exceptions are increases in precipitation rates in North American coastal regions, especially in coastal southern Alaska (≤ 2300 mm a-1) and the US Pacific Northwest (≤ 1700 mm a-1). Further exceptions are precipitation increases in low-altitude regions immediately east of the Peruvian Andes (≤ 1800 mm a-1), central Bolivia (≤ 1000 mm a-1), most of Chile (≤ 1000 mm a-1), and northeastern India (≤ 1900 mm a-1). Regions of notable precipitation decrease are northern Brazil (≤ 1700 mm a-1), southernmost Chile and Argentina (≤1900 mm a-1), coastal south Peru (≤ 700 mm a-1), central India (≤ 2300 mm a-1), and Nepal (≤ 1600 mm a-1).

Most of the precipitation on land in the PLIO climate is higher than that in the PI climate. Precipitation is enhanced by ca. 100–200 mm a-1 in most of the Atacama Desert, by ≤ 1700 mm a-1 south of the Himalayan orogen, and by ≤ 1400 mm a-1 in tropical South America. Precipitation significantly decreases in central Peru (≤ 2600 mm), southernmost Chile (≤ 2600 mm), and from eastern Nepal to northernmost northeastern India (≤ 250 0mm).

In addition to the global changes described above, the PLIO to PI regional climatology changes substantially in the four investigated regions of South Asia (Sect. 3.3.1), the Andes (Sect. 3.3.2), southern Alaska (Sect. 3.3.3), and the Cascade Range (Sect. 3.3.4). Each climate cluster defines a separate distinct climate that is characterised by the mean values of the different climate variables used in the analysis. The clusters are calculated by taking the arithmetic means of all the values (climatic means) calculated for the grid boxes within each region. The regional climates are referred to by their cluster number C1, C2, …, Ck, where k is the number of clusters specified for the region. The clusters for specific palaeoclimates are mentioned in the text as Ci[t], where i corresponds to the cluster number (i= 1, …, k) and t to the simulation time period (t= PI, MH, LGM, PLIO). The descriptions first highlight the similarities and then the differences in regional climate. The cluster means of seasonal near-surface temperature amplitude and seasonal precipitation amplitude are referred to as temperature and precipitation amplitude. The median, 25th percentile, 75th percentile, minimum, and maximum values for annual mean precipitation are referred to as Pmd, P25, P75, Pmin, and Pmax, respectively. Likewise, the same statistics for temperature are referred to as Tmd, T25, T75, Tmin, and Tmax. These are presented as box plots of climate variables in different time periods. When the character of a climate cluster is described as “high”, “moderate”, and “low”, the climatic attribute's values are described relative to the value range of the specific region in time; thus high PLIO precipitation rates may be higher than high LGM precipitation rates. The character is presented in a raster plot to allow compact visual representation of it. The actual mean values for each variable in every time slice and region-specific cluster are included in tables in the Supplement.

This section describes the major changes in regional climatology and highlights their possible implications on erosion rates.

Southern Alaska is subdivided into two wetter and warmer clusters in the south and two drier, colder clusters in the north. The latter are characterised by increased seasonal temperature variability due to being located at higher latitudes (Fig. 12). The different Equator-to-pole temperature gradients for LGM and PLIO may affect the intensity of the Pacific–North American teleconnection (PNA; Barnston and Livzey, 1987), which has significant influence on temperatures and precipitation, especially in southeastern Alaska, and may in turn result in changes in regional precipitation and temperature patterns and thus on glacier mass balance. Changes in the Pacific Decadal Oscillation, which is related to the PNA pattern, has previously been connected to differences in late Holocene precipitation (Barron and Anderson, 2011). While this climate cluster pattern appears to be a robust feature for the considered climate states, and hence over the recent geologic history, the LGM sets itself apart from PI and MH climates by generally lower precipitation (20–40 mm) and lower temperatures (3–5 ∘C; Figs. 10, 11), which may favour frost-driven weathering during glacial climate states (e.g. Andersen et al., 2015; Marshall et al., 2015) in unglaciated areas, whereas glacial processes would have dominated most of this region as it was covered by ice. Simulation PLIO is distinguished by temperatures that exceed PI and MH conditions by ca. 2 ∘C and by larger temperature and precipitation value ranges, possibly modifying temperature- and precipitation-dependent erosional processes in the region of southern Alaska.

In all time slices, the geographic climate patterns, based on the cluster analysis (Fig. 15), represent an increase in the degree of continentality from the wetter coastal climates to the further inland climates with greater seasonal temperature amplitude and lower precipitation and precipitation amplitude (Fig. 15e–h). The most notable difference between the time slices is the strong cooling during the LGM, when temperatures are ca. 13 ∘C (Figs. 13, 14) below those of other time periods. Given that the entire investigated region was covered by ice (Fig. 13), we can assume a shift to glacially dominated processes.

The predicted precipitation differences reported in this study were compared with observed (proxy record) palaeoprecipitation change. Proxy-based precipitation reconstructions for the MH and LGM are presented for South Asia and South America for the purpose of assessing ECHAM5 model performance and for identifying inconsistencies among neighbouring proxy data. Due to the repeated glaciations, detailed terrestrial proxy records for the time slices investigated here are not available, to the best of our knowledge, for the Alaskan and Pacific NW USA studies. Although marine records and records of glacier extent are available in these regions, the results from them do not explicitly provide estimates of wetter–drier or colder–warmer conditions that can be spatially compared to the simulation estimates. For these two areas with no available records, the ECHAM5-predicted results therefore provide predictions from which future studies can formulate testable hypotheses to evaluate.

The palaeoclimate changes in terrestrial proxy records compiled here are reported as “wetter than today”, “drier than today”, or “the same as today” for each of the study locations and plotted on top of the simulation-based difference maps as upward-facing blue triangles, downward-facing red triangles, and grey circles, respectively (Figs. 16, 17). The numbers listed next to those indicators are the ID numbers assigned to the studies compiled for this comparison and are associated with a citation provided in the figure captions.

In South Asia, 14 out of 26 results from local studies agree with the model-predicted precipitation changes for the MH. The model seems able to reproduce the predominantly wetter conditions on much of the Tibetan Plateau, but predicts slightly drier conditions north of Chengdu, which is not reflected in local reconstructions. The modest mismatch between ECHAM5-predicted and proxy-based MH climate change in South Asia was also documented by Li et al. (2017), whose simulations were conducted at a coarser (T106) resolution. Despite these model–proxy differences, we note that there are significant discrepancies among the proxy data themselves in neighbouring locations in the MH, highlighting caution in relying solely upon these data for regional palaeoclimate reconstructions. These differences could result from either poor age constraints on the reported values or systematic errors in the transfer functions used to convert proxy measurements to palaeoclimate conditions. The widespread drier conditions on the Tibetan Plateau and immediately north of Laos are confirmed by seven out of seven of the palaeoprecipitation reconstructions. Of the reconstructed precipitation changes, 23 out of 39 agree with model predictions for South America during the MH. The model-predicted wetter conditions in the central Atacama Desert, as well as the drier conditions northwest of Santiago are confirmed by most of the reconstructions. The wetter conditions in southernmost Peru and the border to Bolivia and Chile cannot be confirmed by local studies. Of the precipitation reconstructions, 11 out of 17 for the LGM are in agreement with model predictions. These include wetter conditions in most of Chile. The most notable disagreement can be seen in northeastern Chile at the border to Argentina and Bolivia, where model-predicted wetter conditions are not confirmed by reported reconstructions from local sites.

Model performance is, in general, higher for the LGM than for the MH and overall satisfactory given that it cannot be expected to resolve sub-grid-scale differences in reported palaeoprecipitation reconstructions. However, as mentioned above, it should be noted that some location (MH of South Asia, and MH of norther Chile) discrepancies exist among neighbouring proxy samples and highlight the need for caution in how these data are interpreted. Other potential sources of error resulting in disagreement of simulated and proxy-based precipitation estimates are the model's shortcomings in simulating orographic precipitation at higher resolutions, and uncertainties in palaeoclimate reconstructions at the local sites. In summary, although some differences are evident in both the model–proxy data comparison and among neighbouring proxy data themselves, the comparison above highlights an overall good agreement between the model and data for the South Asia and South American study areas. Thus, although future advances in GCM model parameterisations and new or improved palaeoclimate proxy techniques are likely, the palaeoclimate changes documented here are found to be in general robust and provide a useful framework for future studies investigating how these predicted changes in palaeoclimate impact denudation.