All simulations were performed with the Land surface Processes and eXchanges (LPX-Bern) dynamic global vegetation model (DGVM) version 1.4 . The model setup is mostly identical to , which is briefly described below. LPX-Bern includes an interactive carbon, water, and nitrogen cycle and simulates dynamic vegetation composition with plant functional types (PFTs), which compete for water, light, and nutrients . The implementation of permafrost and peatlands as long-term carbon stores is based on the LPJ-WHyMe model with the addition of dynamic peatland area .

Peatland vegetation is represented by five peat plant functional types (PFTs): Sphagnum and flood-tolerant graminoids as indicative mostly for high-latitude peatlands, and flood-tolerant tropical evergreen, deciduous tree PFTs, and a flood-tolerant C4 type grass as indicative mostly for tropical peatlands . Carbon cycling in peat soils is based on the distinction between a lower, fully water saturated slow overturning pool (catotelm; from 0.3 to 2 m of the soil column) and an upper, fast-overturning pool (acrotelm; upper 0.3 m of the soil column) with fluctuating water table position (WTP) . Decay rates are modulated by temperature in the catotelm and by temperature and WTP in the acrotelm . The size and sign of the carbon flux between acrotelm and catotelm are determined by the acrotelm carbon balance. Methane emissions from peatlands are simulated but not part of the analysis in this study.

The area fraction covered by peat in a given grid cell is determined dynamically with the DYPTOP module (Dynamical Peatland Model Based on TOPMODEL) . The TOPMODEL approach is used to predict the monthly inundated area fraction given sub-grid-scale topographic information and mean grid cell WTP, averaged over all land classes. The area potentially available for peatlands is then determined by inundation persistency. Peatlands expand or shrink towards a changing potential extent with a rate of 1% of their current grid cell fraction per year. The grid cell fraction lost during peatland retreat is treated as a separate land class for former peatlands. It inherits the carbon stocks of the shrinking peatland and is subsequently treated in the same way as the mineral soils regarding vegetation, hydrology, and carbon cycling. Growing active peatlands first expand on former peatlands inheriting the remaining carbon there. This treatment prevents carbon dissolution into mineral soils due to fluctuations in peatland area.

Different to we also consider changing land-use area in our simulations. Land-use area in the model is represented by three specific land classes: pasture, cropland, and urban . Pastures and croplands have specific vegetation represented by two herbaceous PFTs each. Changes in land-use area are treated as net changes, where all growing land classes proportionally inherit the carbon, water, and nutrients of all shrinking land classes. A more complex implementation, which considers gross changes, exists but is not compatible with the peatland module used here . In the absence of gross change information, three assumptions were made. (1) Changes within the three land-use classes that do not affect the total land-use area are assumed as shifts between land-use types (e.g., shift from pasture to cropland). (2) Increases in total land-use area reduce all other land classes proportionally, including peatlands. (3) Peatland area that is converted to land-use area cannot be reclaimed by expanding peatlands at a later stage. These assumptions are simplifications that fail especially in areas where peatlands are preferentially targeted for land-use conversion, such as in Indonesia , or are subject to restoration efforts after conversion e.g.,. However, given the technical restrictions and the lack of detailed worldwide information about gross land-use changes on peatlands, we think this simplified approach is the most robust.

The above-described representation of peatlands in the LPX is a simplification in many respects. The absence of local processes and information such as lateral water flow, local soil features, or influence of animals by grazing and river damming can limit the ability of the TOPMODEL approach to predict peatlands on a regional to local scale. Further, direct human-caused influences such as land use, drainage, or peat mining are only considered in a strongly simplified way. The lack of a distinction and transition between different peatland types like fens, bogs, blanket bogs, or marshes neglects possible differences in the constraints on their formation and evolution. The treatment of acrotelm and catotelm as single carbon pools, and the absence of strong disturbances such as peat fires, constitute limits on the comparability of the model results to peat core carbon profiles. This simplified representation, nevertheless, has been shown to reproduce peatland area and carbon accumulation well within the observational constraints while using a minimal set of free parameters.

Peat carbon can be present not only in soils of active but also in the soils of former peatlands. Peat may be preserved during peatland conversion and form distinct organic soil layers on non-peatland areas . In the model, subsequent land classes inherit the soil carbon from former peatlands, including peatlands converted to land-use areas. Yet, this peat carbon is mixed within the model's soil pools and cannot be directly distinguished from carbon transferred to soils from more recently established vegetation. It is, however, possible to track peat carbon that at one point was sequestered in the catotelm of active peatlands through the soil pools of other land classes using post processing. For this, transient model output for peatland area changes, the decay rates of slow overturning pools, and the carbon input into the catotelm of active peatlands is needed. Area changes are used to transfer carbon between active peatlands, former peatlands, land-use areas, and natural vegetation classes. Transient decay rates are used to decay the carbon in the respective pools. Carbon is thus tracked from its entry into the catotelm of an actively accumulating peatland until its decay there or in a former peatland or land-use area. This approach cannot take account of the acrotelm carbon. However, acrotelm carbon constitutes only a small part of simulated total peatland carbon (5 % at 1975), and we can assume that this carbon at the peat surface is quickly respired after peatland transformation. For the analysis, we refer to two different variables related to peat carbon: (1) peatland carbon, which refers to the carbon stored in the acrotelm and catotelm pools of active peatlands, and (2) total peat carbon, which is calculated in post-processing and represents all carbon in the catotelm of active peatlands and organic, not-yet-decomposed carbon that was at some point sequestered into a catotelm on peatlands transformed to land-use areas and other former peatlands. After ecosystem transformation, depending on the transition and the conditions thereafter, former peatlands can see a fast collapse or erosion of carbon stocks as well as buried peat carbon layers preserved for millennia . The two carbon variables can be interpreted as two bounding cases to the fate of peat carbon in former peatlands. Changes in the variable peatland carbon can represent a fast emission bounding case where peatland carbon is lost immediately after ecosystem or land-use transformation. The slow emission bounding case, with peat carbon decaying in former peatlands over a long timescale, can be represented by changes in the variable total peat carbon. The true fate of peat carbon in former peatlands in most cases will lie somewhere in between these worst- and best-case scenarios.

The simulations presented here are a direct continuation of a transient simulation from the Last Glacial Maximum (LGM) to the present, which was discussed in detail in . This enables future projections starting from a truly transient spinup, including all potential legacy effects of the past 22 000 years. The LGM simulation was run with a model resolution of 2.5∘ latitude × 3.75∘ longitude and was forced with CO2 and temperature and precipitation fields. Temperature and precipitation anomalies were taken from the transient CCSM3 run TraCE21k . The TraCE21k anomalies were imposed on the CRU TS 3.1 base climate from 1960 to 1990. Interannual variability thus came from TraCE21k. Temperature anomalies were calculated as absolute and precipitation anomalies as relative values.

The resolution of the LGM simulation was adopted for the future simulations. This ensures a truly seamless transition between the simulations, without unpredictable effects of downscaling on peatland dynamics.

In the original LGM simulation, land use was not considered as the focus of the study was on the natural development and evolution of peatlands since the LGM. To integrate a transient history of land use, the simulation was restarted in the year 1500 with subsequent transient land-use forcing and otherwise unchanged boundary conditions.

In the year 1975, the midpoint of the base climate period, forcing transitions from TraCE21k to CMIP6 climate anomalies (temperature, precipitation, and cloud cover) see Fig. , whereas the base climate remains unchanged. From this point on, simulations are done for each model of a 10-member climate model ensemble (see Sect. ). Short historical simulations from 1975 to 2014 bridge the gap between the LGM simulation and the start of the CMIP6 scenarios with anomalies taken from the CMIP6 historical simulation of the climate model ensemble.

Simulations corresponding to three different CMIP6 scenarios start from the year 2015. One strong-mitigation scenario (SSP1-2.6), one middle-of-the-road scenario (SSP2-4.5) and one high-emission scenario (SSP5-8.5) were selected to represent the scenario range. The standard CMIP6 scenarios end in the year 2100. To investigate the delayed long-term responses of peatlands, the forcing is extended into the future with a detrended version of the last 30 years of each time series repeated over almost 5000 years until the year 7000. The trend correction was done per grid cell and month and with respect to the end of the time series. Scenario CO2 forcing was adopted from . Land-use forcing is taken from the Land-Use Harmonization (LUH2) project see Fig. g–i,.

CMIP6 also includes extended versions of the scenarios SSP1-2.6 and SSP5-8.5 that range until 2300. At the time of this study, however, only three climate models had provided output for these extended scenarios. Climate projections of these three models alone are not representative of the full CMIP6 scenario. They were, however, included in the 10-member climate model ensemble used here (see Sect.  and Fig. S1) and additional simulations with transient climate and CO2 forcing until 2300 were performed to compare results to the standard simulations.

To disentangle future changes in peatlands that are induced by changes in climate, CO2, and land use up to 2014 from those induced by future changes in these drivers, we performed an additional simulation with constant boundary conditions at 2014 levels for each ensemble member. Here climate forcing was extended with a detrended version of the last 30 years of the historical forcing. Similarly, a control simulation was performed with constant boundary conditions after 1500 to show the undisturbed model state. These simulations reveal the committed changes in peatland area and carbon induced by the deglacial changes prior to the preindustrial state at 1500 and the changes over the historical period until 2014 respectively.

To determine the different driver contributions to the changes in peatland variables, additional factorial simulations were performed for all scenarios and climate model ensemble members. For each standard simulation, there are five factorial simulations with one of the five transient forcings (temperature, precipitation, cloud cover, CO2, and land use) kept constant at 2014.

The driver contribution to the anomaly of peatland variables was determined as the difference between the standard run anomaly and the anomaly in the respective factorial run. The contribution from already committed changes due to past climate and land-use change was determined as the anomalies in the simulations with overall constant forcing after 2014 (see Sect. ). The residual of the difference between the sum of all contributions and the standard run anomaly was identified as the contributions from non-linear interactions and other factors not considered in the analysis. Cloud cover was found to have only a minimal effect on the considered peat variables in the LPX-Bern, and thus for further analysis its contribution was added to the other/non-linear category.

As a second step, driver contributions were classified as driving contributions (same sign as peatland variable anomaly) and dampening contributions (opposite sign as peatland variable anomaly) and re-normalized respectively. Figures show only the driving contributions of the respective positive and negative peatland variable anomalies.

We chose a subset of 10 climate models out of a CMIP6 ensemble of 22 models (see Figs. a–f and S1) that at the time (June 2020) provided monthly output for all necessary forcing variables – precipitation, near-surface temperature, and cloud cover – and for all considered experiments – historical, SSP1-2.6, SSP2-4.5, and SSP5-8.5. Climate model output was downloaded from the earth system grid database. One additional model, the Community Integrated Earth System Model (CIESM), had to be excluded from the CMIP6 ensemble as it showed a discontinuity in the precipitation data between the historical and the scenario simulations.

Three models, IPSL-CM6A-LR, MRI-ESM2-0, and CanESM5, were included in the subset a priori as they were the only ones that also provided output for both of the extended SSP1-2.6 and SSP5-8.5 scenarios (see Sect. ). The other seven were chosen for the ensemble subset to optimally represent the full CMIP6 ensemble as a whole. The optimization targets of ensemble total range, interquartile range, and median, were defined with respect to the anomalies (from 1961–1990 to 2071–2100) in precipitation and temperature as the most important forcings to the LPX-Bern. The optimization was inspired by . A total of 2000 randomly drawn subsets were ranked according to the distance of the subset to the targets with normalized scores calculated and averaged over all individual grid cells, months, scenarios, and variables. The rating of the best-performing subsets was further improved by a careful handpicked combination, resulting in the final subset including the climate models referenced in Table .

This subset performs best when calculated both over the total land area and over the simulated peatland area alone. Over land, temperature anomaly total range, interquartile range, and median differ between the ensemble subset and the full ensemble by 0.25, -0.18, and 0.02 ∘C respectively, with larger distance at higher-emission scenarios. Averaged over the simulated peatland area at 1975, the distances are 0.28, -0.31, and 0.03 ∘C respectively. For precipitation, anomaly total range, interquartile range, and median differ between full ensemble and ensemble subset by 9.7, 0.64, and -0.15 mm over all land area and 7.3, 0.54, and -0.49 mm over the peatland area respectively. The optimization procedure thus yielded an ensemble subset representative of the full CMIP6 ensemble, although, given the number of possible combinations, optimization could be improved further with further sampling.

There are still considerable uncertainties connected to estimates of the global area covered by peatlands and the amount of organic carbon stored within them. Estimates for northern peatland area, using various methods ranging from inventory based to machine learning, lie between 2.4 and 4.0 million square kilometers (Mkm2) . For tropical peatlands, which are still much less studied than northern peatlands, peatland area estimates have increased in recent years following the discovery of large new peatland complexes, such as in the Congo Basin , and due to new methodologies trying to account for potentially undiscovered peatlands . Earlier estimates of tropical peatland area thus range from 0.37 to 0.44 Mkm2 and more recent estimates from 1.0 to 1.7 Mkm2 . Peatland areas simulated by LPX in the year 1975, the end of the transient simulation from the LGM, are within the range of literature estimates with a global, northern (>30∘ N), and tropical (30∘ S to 30∘ N) peatland area of 3.8, 2.8, and 1.0 Mkm2 respectively. Global peatland area is shifted more towards the tropics as in most estimates. However, most major peatland complexes seen in global peatland maps, e.g., PEATMAP see Fig. , are captured well. Major regional differences exist in Africa, where LPX-Bern fails to simulate the large Congo Basin peatland complex. Peatland area is also underestimated in northern Europe. In North America the model overestimates peatland area in Alaska and Quebec and underestimates peatland extent in western Canada.

Estimates of global peatland carbon are directly dependent on peatland area estimates and thus also come with a large uncertainty range. Northern peatlands have been estimated to store 270 to 604 gigatons of carbon (GtC), using various methods and area estimates (see , and , for a review). For tropical peatlands, estimates of organic carbon storage range from 44 to 92 GtC in earlier estimates and increase as a result of larger assumed areas in recent estimates from 70 to 288 GtC . In the year 1975, peatlands simulated by the LPX-Bern have accumulated about 441 GtC of soil organic carbon globally. From this, northern and tropical peatlands make up 319 and 121 GtC respectively. Simulated carbon stocks thus lie within the literature estimates but again with a distribution shifted more towards the tropics than most estimates suggest.

Throughout the 22 000 years of the transient simulation up to the year 1975, peatland area was highly dynamic with today's peatlands gradually expanding but also large paleo-peatlands vanishing over time . There, similarly as described in Sect. , the carbon from former peatlands was tracked through subsequent land classes until its decay. At 1975 LPX-Bern gives a total of 195 GtC of peat carbon leftover from former peatlands on land, with 165 GtC in northern latitudes and 28 GtC in the tropics. Total peat carbon is thus simulated to be 612 GtC. Very little is known about the amount and location of peat leftover and buried from former peatlands, although various deposits have been found . Peat carbon lost due to past or future sea-level rise is not considered in this study.

Unlike the original LGM simulation in , here land use and land-use change is considered since the year 1500, with land-use areas being able to expand onto peatlands (see Sect. ). Up to 1975 about 0.42 Mkm2 of peatland area and 50 GtC of peatland carbon are lost from active peatlands due to land-use change in the simulation. Carbon loss is reduced to about 5 GtC when considering total peat carbon, which does not assume an immediate loss but a slow decay in former peatlands. estimate that about 0.51 Mkm2 of peatland area and 22 ± 5 GtC of peatland carbon was lost globally from 1850 to 2015 due to drainage and land-use conversion of peatlands, which is in rough agreement with the simulated values (see also Sect. ).

A factorial analysis was used to attribute the positive and negative changes in peatland variables to individual forcing drivers (see Sect.  ). Figure  shows the calculated mean driver contributions to the global gross positive and negative anomalies in peatland area, peatland carbon, and total peat carbon.

Increases in peatland area up to 2300, both in the high and low latitudes, are driven mostly by committed changes (constant 2014 conditions) and an increase in regional precipitation. In northern permafrost regions, this is further strengthened given strong mitigation and moderately rising temperatures resulting in longer growing seasons and larger water retention (Fig. ). Suggested pathways of permafrost peatlands after thaw are still debated but include rapid degradation , collapse followed by long-term recovery , and increased carbon accumulation , depending on multiple factors such as thaw velocity and local hydrology.

Simulated peatland area losses are driven mostly by committed changes and increasing temperatures. Higher temperatures lead to an increase in evapotranspiration, especially in boreal peatlands , and thus a decrease in the regional water balance which is not compensated for despite a potential concurrent increase in annual precipitation. This corresponds to the already observed decade-to-century-long drying trends in northern Europe and eastern Canada peatlands , regions of large simulated committed area loss, which is found to result in negative effects on carbon accumulation rates and strong trends of woody encroachment. These trends are expected to continue and amplify under future climate change.

In the LPX-Bern, a decreasing water balance can lead to a positive feedback on the retreating water table. A long-term drawdown of the mean grid cell water table leads to a reduction in peatland area, which in turn reduces the mean grid cell water table further. In some cases, this can lead to a much larger reduction in peatland area than would be the result of the initial disturbance. In reality a lot of complex and often still poorly understood both positive and negative hydrological feedbacks control the peatland water table position in response to disturbances . The simplified structure of DGVMs like the LPX-Bern cannot mirror these complex interactions, and thus this strong internal feedback needs to be interpreted with caution. However, although peatlands, in some cases, have been found to be relatively resilient with respect to limited disturbances , there could be possible tipping points that could lead to fast vegetation and ecosystem transitions under strong persisting disturbance .

Peatland carbon and total peat carbon dynamics up to 2300 are dominated by committed and temperature-driven losses. The decline in peatland area directly reduces global peatland carbon and indirectly affects the total peat carbon balance by reducing overall accumulation. At the same time, global decay in active as well as in former peatlands is increased by the higher temperatures. Peatland NEP in the northern high latitudes is also driven in large parts by committed changes and increasing temperatures which can both increase or decrease NEP given the balance between respiration and productivity and their effect on permafrost (Figs.  and ). Increasing CO2, precipitation, and non-linear interactions between the drivers have strong positive effects on NEP in East Asia. In the tropics, the negative trend in NEP is mostly driven by the higher temperatures and non-linear effects.

Up to 2300, land-use change and atmospheric CO2 have a comparatively small impact on the global scale. Increasing CO2 concentrations have a positive effect on carbon accumulation which is progressively larger with increasing emission scenarios. This also translates into a moderate peatland area gain at 2300. Effects of land-use change are negative for all peatland variables but remain small. One reason for this small impact might be that only net land-use area increases are considered to affect peatlands in the model. The global net increase in land-use area, however, is much smaller in the future scenarios, than for the historical period (Sect. ). Especially in regions where peatlands are directly targeted for land-use conversion, such as in Indonesia , our approach might significantly underestimate the negative effect of land-use change.

The late expansion after 2300 is driven by peatlands newly establishing in model grid cells with no previous peatland presence. In some grid cells, the historical and future climate change and atmospheric CO2 rise lead to the fulfillment of criteria for peatland establishment, targeting the peatland water and carbon balance. The more the boundary conditions change under future scenarios, the more grid cells, especially in the tropics, become able to support peatlands (see Sect. ). Initiation of new peatlands in the tropics is mostly driven by CO2 fertilization nudging the carbon balance over the initiation thresholds (Fig. ). In northeastern Asia, the temperature rise is the most prominent initiation driver, whereas in northwestern Canada mostly precipitation increases drive the initiation of new peatlands. In the model, newly established peatlands start from a small seed, and their growth is restricted to 1 % of their size per year. They thus reach noticeable size only centuries after their initial establishment. Outside of the model world, the speed of lateral expansion of growing peatlands depends on multiple factors, including local topography, hydrology, and peatland type . Topography can constrain lateral expansion velocities. Depending on terrain slopes, peat accumulation can be limited to a small area or depression for centuries to millennia until the peat column grows tall enough or expand quickly over a flat plain . This heterogeneity and complexity in lateral expansion of newly established peatlands is not represented by the model used here. The magnitude and timing of the simulated late expansion should therefore be taken with care. However, the results suggest that historical and future climate change might create the potential for newly forming peatlands in regions where conditions have been mostly unsuitable before.

The spread between simulations forced with climate anomalies from the different CMIP6 climate models indicates a large climate-anomaly-related uncertainty in simulated peatland variables. This is in line with previous studies that also found a large uncertainty propagation from climate variables to peatland and carbon cycle variables in general .

The magnitude of uncertainties is regionally different, with large uncertainties in the northern high latitudes for peatland area (Fig. S3) and total peat carbon (Fig. S4) and in the mid-latitudes for peatland NEP (Fig. S5). The climate variables driving the uncertainty depend on the region and peatland variable in question. Linear regressions for each grid cell were used to investigate how the differences between the model climate anomalies translate to the simulated peatland variables. Differences in northern-high-latitude peatland area between simulations were found to be dominantly a factor of climate model temperature. Warmer anomalies resulted in less peatland area in most grid cells, except for northeastern Asia, where warmer temperatures facilitate peat expansion in some grid cells (Fig. S3). In the tropics, the difference in precipitation is the best predictor for most grid cells, with anomalies from wetter models resulting in larger peatlands. Total peat carbon, determined by the balance between total accumulation and total decay of peat carbon, shows a similar regional pattern (Fig. S4). Northern-high- and mid-latitude peat carbon is reduced with higher temperature anomalies as the area for accumulation declines and heterotrophic respiration increases. In the tropics precipitation remains the dominant predictor, increasing the accumulation area and limiting respiration. For peatland NEP, precipitation minus evapotranspiration, as a measure of the moisture balance, resulted in a larger number of grid cells with significant (regression p value <0.05) results compared to temperature or precipitation alone (Fig. S5). Climate models resulting in a more positive water balance mostly also resulted in a higher peatland NEP due to the controls of peatland water table depth on both productivity and respiration. In permafrost regions, also temperature on its own is a strong positive factor for simulated peatland NEP.

The results show that the differences in peatland responses to different climate forcings can be explained mostly by the same drivers and mechanisms as the transient changes (discussed in Sect. ). They also reveal the large dependence of peatland and carbon cycle projections on key properties of climate models. Further constraining model climate sensitivity is thus essential to reduce uncertainty in carbon cycle projections. Here the climate model ensemble subset was selected to best represent the full CMIP6 ensemble and with this the fullest possible range of projections. However, model performance compared to different targets is highly variable , and different ensemble subset selections or weighted ensemble medians might be preferable in future work, depending on the focus.

The simulations presented here are also subject to other large but less quantifiable uncertainties. Keeping the forcing constant after 2100 or 2300 is an idealization with true climate dynamics depending on highly uncertain factors such as future social and economic dynamics and potential tipping points. Structural and parameter uncertainties, not only in the peat module, but through all components of the model, are unavoidable in simplified global models such as the LPX-Bern, especially on regional and local scales. Implementation of peatlands in DGVMs is still in its early stages, and comprehensive model comparison and structural uncertainty evaluation are still mostly lacking. With the inclusion of peatlands into more and more DGVMs and earth system models, comparative studies might identify the most promising model developments and thus pave the way for more robust peatland and carbon cycle projections.