GVM simulation results were obtained from the ISIMIP repository (https://data.isimip.org, last access: 14 November 2025, see Table S1 for summary of all data sets used for the analysis). Specifically, we used simulation results from ISIMIP2b simulations for the climate change scenarios RCP2.6, RCP6.0, and RCP8.5, simulated with climate forcings from IPSL-CM5A-LR or HADGEM2-ES Earth System Models ESMs,. Both climate forcing data sets were bias corrected with EWEMBI observational data . ISIMIP2b simulation runs include different combinations for direct anthropogenic impacts and atmospheric CO₂ (increasing or fixed). Here, we used the scenario where land use, nitrogen deposition and fertilizer input was included but fixed at the levels of the year 2005 (“2005soc” scenario in ISIMIP repository) and where CO₂ increased according to the respective RCP scenario (“co2” scenario in ISIMIP repository). We selected the combination of IPSL-CM5A-LR, “2005soc” and “co2”, because for this combination, results were available for five different GVMs for RCP6.0. This combination of scenarios had the highest number of GVMs available, ensuring model comparability. Model results were available for LPJ-GUESS , ORCHIDEE , ORCHIDEE-DGVM , CLM4.5 , and CARAIB . For RCP2.6 and RCP8.5, we used HADGEM2-ES results for CLM4.5, as results for IPSL-CM5A-LR were not available in the repository. For RCP8.5, results were only available for LPJ-GUESS, ORCHIDEE and CLM4.5. As we did not run model simulations for this study but only used available model results, we do not provide model descriptions or model comparisons and refer to the key references for the different models (Supplement Sect. S1, Tables S2 and S3 for temperature-sensitive parameters and bioclimatic limits in the GVMs).

The target variable for biome classification was leaf area index (LAI), because this variable has historically been used for biome classification e.g.,. Using LAI ensures consistency with previous studies. The GVMs simulate state variables such as carbon pools, height or LAI separately for each PFT in response to environmental conditions, competition and disturbances such as fire . The cover of different PFTs within grid cells can be static and informed by land cover maps (ORCHIDEE, CLM4.5) or change dynamically in response to biotic and abiotic factors (LPJ-GUESS, ORCHIDEE-DGVM, CARAIB). For our analyses, we only used PFTs representing natural vegetation while bare ground and anthropogenic PFTs were ignored (see Supplement Sect. S1 for details). Annual LAI data were averaged for the 15-year periods 2006–2020 and 2085–2099, representing early and late conditions within the future simulations of ISIMIP2b that were provided for the period 2006–2099 for each climate scenario. Climate conditions in different RCPs do not differ strongly for the early period (2006–2020) but diverge after that period. Therefore, we hereafter denote the early period as “current” irrespective of the RCP scenario, and the late period as “future”. LAI data in the ISIMIP repository did not account for cover fractions of different PFTs within grid cells. Therefore, we processed LAI data in two steps. First, we multiplied PFT-specific LAI with the cover fraction of the respective PFT. Then, we scaled the LAI of natural PFTs with the inverse sum of the cover fractions of all natural PFTs. The second scaling step removes anthropogenic PFTs and assumes that the entire grid cell is only covered by natural PFTs (see Supplement Sect. S1 natural/anthropogenic PFTs). This scaling is reasonable, as we only consider natural biome types in our biome classification (see section “Biome classification”), while all anthropogenic PFTs and land cover types are excluded. The data were available at 0.5° spatial resolution.

We used 31 biome maps compiled by , hereafter called F31 biome maps. Those maps were published between 1964 with adjustments and revisions afterwards and 2020 , and represent historic biome patterns observed before the publication date. Biome states may have shifted since the publication date. Yet, as the time period represented by the biome maps is not always clearly defined and as we use 15-year periods for GVM results, we ignored potential biome shifts after the publication of biome maps and compared all maps to the GVM results for current conditions (2006–2020). The approach is consistent across all observation-based biome maps, even though it may introduce minor uncertainties due to mismatches in the time periods. The biome maps were derived from different quantities and by different methods, such as biogeographic mapping of species, bioclimatic zonation or classification using remote sensing data. Hence, the biome maps represent different functional and structural features of vegetation. The maps differ substantially with respect to the number of biomes included, the definition of different biome types, and accordingly their spatial distributions. As GVM results from the ISIMIP2b repository were provided at 0.5° spatial resolution (see previous paragraph), we aggregated the biome maps to the same 0.5° resolution using the nearest neighbor method. In our analysis, we created random forest models for each combination of the 31 observation based biome maps and the GVMs and did not aggregate them into consensus biome or GVM maps. This approach is necessary because the biome maps and GVMs use incompatible biome and PFT types, respectively. Aggregation into consensus maps would itself require a series of subjective decisions and would introduce additional uncertainty. Further, our approach allows us to explicitly quantify uncertainties related to biome classification.

In addition to GVM results, we used observation-based PFT maps for our analysis: the European Space Agency (ESA) Climate Change initiative (CCI) product and the product. The products include different sets of PFTs (Supplement Sect. S1). Those PFT maps were derived from different remote sensing products. developed a cross-walking scheme to match PFT distributions with land cover classes from the CCI Multi-Resolution Land Characteristics (MRLC) product, supported by products describing for example surface water, built-up areas or tree canopy cover and height. used classification and data integration approaches to create a consensus map based on reflectance data from GlobCover Land Cover (GlobCover), Moderate Resolution Imaging Spectroradiometer (MODIS2005), Global Land Cover 2000 (GLC2000) and IGBP Data (DISCover). Even though those products are observation-based, they are expert-based and subjective and affected by uncertainties . Classification with those products serves as reference for the evaluation of GVM results. Those maps are only available for current conditions, while future changes of PFT maps are not available. Both products were aggregated to the 0.5° spatial resolution of the GVM results using bilinear interpolation.

We used random forests to classify PFT information from GVMs (based on LAI) or remote sensing (cover fraction) into the F31 observation-based biomes, that is, we applied supervised classification. Random forests were fitted using the “randomForest” R package version 4.7-1.2,. For each combination of one of the F31 biome maps, modeled LAI or observation-based PFT cover map, and RCP scenario, one random forest model was created. Simulated LAI for current conditions or cover fractions were used as explanatory variables, the biome types of the respective biome maps were used as response variables, where all biome types were classified in one random forest model. We used the “randomForest” default settings. For each random forest, 2500 trees were created. Model performance was evaluated using out-of-bag (OOB) error estimation for each tree, therefore we did not split data into training and testing subsets. Using the random forests, we then predicted biome patterns for current conditions for each GVM or PFT map and for each of the F31 biome maps. Overall, we created 62 biome maps for PFT products (31 biome maps, 2 PFT products), 155 maps for five GVMs for RCP2.6 and RCP6.0, and 93 maps for three GVMs for RCP8.5. We further applied the random forest models to the PFT-specific LAI simulated by GVMs for future (period 2085–2099) climate and CO₂ conditions of the respective GVM, to obtain biome patterns for future conditions. This was only possible for the five GVMs, as future projections for the PFT cover products were not available. Overall, we created 403 biome maps for different combinations of GVMs, RCPs and F31 biome maps for both current and future conditions.

For comparisons between observation-based biome maps and the biome patterns obtained from random forest models trained on the respective biome maps, we used the κ-statistics . This quantity allows comparisons between the spatial patterns of categorical variables and considers that agreement can be caused by random effects. Values in the range 0–0.2 indicate slight agreement, 0.2–0.4 fair, 0.4–0.6 moderate, 0.6–0.8 substantial and 0.8–1 almost perfect agreement. We created maps to identify areas where the F31 and respective modeled biome maps agree or disagree, and assessed the relationship between κ values and the number of biomes in the respective biome map. We further calculated κ for each biome type in each F31 biome map and related it to the number of grid cells covered by the respective biome type, to analyze if the performance of the random forest model was related to biome coverage. Therefore, the target biome type was set to one while all other biome types were set to zero, and κ values were then calculated based on the binary maps. We assumed that the entire 0.5° grid cell is covered by only one biome type, both in the F31 maps and the model results. For these analyses of model performance under current climate conditions, we present only results from the RCP6.0 scenario, as vegetation simulations for the current period and agreement with observation-based biome maps are similar.

For the GVM results, we analyzed the susceptibility to biome change under future climate and CO₂ conditions for the entire ensemble of GVM and biome maps and regions of biome stability. Therefore, we created maps indicating biome change or stability between current and future climate conditions for each GVM and for each of the F31 biome maps. Hence, we obtained 155 maps of biome change (5 GVM × 31 biome maps) for RCP2.6 and RCP6.0 and 93 maps for RCP8.5. These maps were then overlaid and the number of models projecting a biome change was counted. Values approaching 155 or 93 indicate that all models consistently predict biome change whereas low values indicate that models consistently predict biome stability. We denote the number of models projecting biome change as the susceptibility of biome change. A high number of GVM-biome-map projecting a biome change in a grid cell is associated with a high susceptibility of biome change, accordingly, a low number is associated with low susceptibility and biome stability. For visualization, we categorized these results, where “low” indicates that 0 % to 20 % of the models predict a biome change (representing low susceptibility and stability), “Medium” indicates 20 % to 40 %, “high” indicates 40 % to 60 % and “very high” more than 60 %. We assessed if the proportion of grid cells affected by biome change was related to the agreement between data and model results (i.e., to the κ value) or to the number of biomes in the respective biome map.

We used a linear regression to test if the number of grid cells affected by biome changes can be explained by the RCP scenario, GVM, F31 biome map, and number of biomes in the biome map. In this analysis, spatial information was aggregated into an average value for each combination of F31 map, GVM and RCP. To identify main drivers of biome change at the grid cell level, we modeled the susceptibility to biome change using a Random Forest (RF) regression using the “ranger” R package version 0.16.0,. Data from RCP2.6, 6.0, and 8.5 scenarios were modeled both individually and in a joint model to capture response patterns across all scenarios. Predictors were current (2006–2020) temperature and precipitation, anomalies between current and future (2085–2099) climate (Figs. S1, S2), and geographic coordinates. To isolate the susceptibility to biome change in different temperature zones, we also used the current temperature zone as predictor (cold <5 °C, temperate 5–18 °C, warm >18 °C). Relative driver influence was assessed via permutation importance, and an interaction model was used to statistically validate sensitivity differences across temperature zones. To isolate responses to temperature change, we created partial dependence plots between biome change susceptibility and temperature change for the three temperature zones.

To exemplify modeled biome distributions and biome change, we analyzed the results for the biome map in more detail. We selected this map as it is commonly used as reference biome map, for example in IPBES or in biome-specific analyses e.g.,. We created maps showing the current and future biome patterns, agreement between modeled and observation-based biome distributions, as well as maps indicating regions where biome shifts occur. To quantify the agreement per biome, we calculated contingency tables for each model indicating the percentages of overlap between observation-based and modeled biomes. In addition, κ values were calculated per biome by setting grid cells covered by the target biome to one and all other biomes to zero, and then calculating the κ value for the binary map. We created Sankey plots to illustrate transitions under biomes between current and future conditions using the “ggsankey” package version 0.0.9999,, and quantified biome coverage and transition rates for current and future biome distributions.

To test if agreement between modeled biomes for current conditions and the biomes is higher for temperature-driven biomes, we first split the biome types into those mainly driven by temperature and those driven by other factors (Table S4). This classification is based on literature and expert-knowledge of the authors, and other authors may use another classification. We did, however, not test the sensitivity of our results to this classification as we used random forest regressions to analyze biome changes for the full model ensemble in response to multiple predictors (see previous section). Then, we calculated the mean κ values for both groups and all GVM and PFT maps. To test if the likelihood of biome change differs between biomes mainly constrained by temperature or other factors, we calculated the percent of grid cells predicted to undergo biome change for both groups. In contrast to the temperature-related analyses of the full ensemble including all 31 biome maps (Sect. 2.3), this approach tests temperature responses indirectly via assumptions on temperature sensitivity of biomes.

To test if biome change is more consistent for biomes with well-defined temperature limits such as chilling requirements or frost tolerance used to define PFTs, we used the coefficient of variation as proxy for consistency. Specifically, we calculated for each GVM and each biome transition the respective percent of grid cells affected by this biome change. Then we calculated mean, standard deviation, and coefficient of variation of these percentages for each biome change across the GVMs. To account only for relevant biome change, we excluded those with less than 5 % coverage under current conditions. Finally, we used a t-test to compare mean coefficient of variation for biomes with well-defined temperature limits and biomes explained by other factors. If biome change for temperature-defined biomes are more consistent, we would expect lower coefficients of variation.

All analyses were conducted using R version 4.4.2,. Spatial data were processed using the “terra” version 1.8-10, package. Figures were created using the “ggplot2” version 3.5.1, package. All R scripts are provided in .

Data-model agreement varied strongly across the 31 biome maps and different LAI or PFT data used for the biome classification (Fig. ). Overall, κ values were higher for biome classification using observation-based PFT data (mean κ=0.95, indicating almost perfect agreement) than for LAI data simulated by GVMs (mean κ=0.77, indicating substantial agreement). Of the F31 biome maps used to inform the biome classification, agreement was highest for the biome map when averaging κ for all GVMs (κ=0.83), for the map for GVMs and PFT maps combined (κ=0.87), and for the biome map using only the observation-based PFT maps for classification (κ=0.98). The lowest overall performance was obtained for the map (κ=0.75). The relation between the κ values and the number of biomes included in the F31 biome maps was weak (Fig. S3, Table S5). The κ values per biome were related to the fraction of grid cells covered by the respective biome, i.e., biomes with large cover had a higher κ values, while biomes with smaller cover showed large variation of κ values ranging between zero and one (Fig. S4).

Regions where observation-based and modeled biomes agreed or disagreed were spatially heterogeneous (Fig. S5). Agreement was high, for example, in Northern Europe, Northern America, and tropical forest regions in Brazil or Southeast Asia (Fig. S5). Agreement was low on the Tibetan plateau, a large part of the sub-tropics in South America and regions bordering the central African rainforests. Those spatial patterns differed between GVMs and observation-based PFT maps (Fig. S5). For instance, CARAIB showed the highest disagreement in northern latitudes and around the Equator, whereas ORCHIDEE-DGVM showed highest disagreement in sub-tropical regions bordering equatorial forests.

The proportion of grid cells projected to be affected by biome change increased from RCP2.6 to RCP6.0 but was lower in RCP8.5 than in RCP6.0. The RCP8.5 scenario lacks CARAIB and ORCHIDEE-DGVM simulations, the two models with the highest rate of biome shift in other RCP scenarios. In addition to the proportion of grid cells, the susceptibility category, i.e., the number of GVM-biome map combinations that predict biome transitions, increased, indicating higher consensus between different models regarding predicted biome change (Fig. ). Thus, the area affected by low susceptibility changed from 53.1 % in RCP2.6 to 32.8 % in RCP6.0 and 40.1 % RCP8.5, while the proportion of grid cells affected by high or very high susceptibility increased from 3.6 % in RCP2.6 to 15.5 % in RCP6.0 and 10.7 % in RCP8.5. Yet, the broad spatial patterns of regions affected by biome change were similar for the different RCPs (Fig. ). For instance, the tropical forests of Africa and Southeast Asia were projected to be stable in all scenarios (grey areas in Fig. ) while northern latitudes or southern Africa showed a susceptibility to biome changes in all RCPs. The spatial patterns of biome change also differed between GVMs (Fig. S6).

For all combinations of the GVMs and F31 biome maps, the proportion of the grid cells affected by biome change ranged between 4 %–45 % for RCP2.6 (155 maps), 7 %–56 % for RCP6.0 (155 maps), and 9 %–53 % for RCP8.5 (93 maps, Figs. , S7). When averaged across all F31 biome maps for each GVM and each RCP, the proportion of grid cells affected by biome change was lowest for CLM4.5 for all RCPs (4.2 %, 6.6 % and 9.3 %) and highest for CARAIB and ORCHIDEE-DGVM for RCP2.6 (44.5 %), CARAIB for RCP6.0 (56.3 %) and LPJ-GUESS for RCP8.5 (52.7 %, with CARAIB and ORCHIDEE-DGVM not being available for RCP8.5). The proportion of grid cells affected by biome shifts was related to data-model agreement (i.e., the κ value of the respective model, Fig. , Table S6). Considering each GVM individually, the proportion decreased when κ is increasing, that is, higher data-model agreement implied a lower rate of biome shifts. When considering all models together, the proportion increased with increasing κ, and the model with the highest overall performance (CARAIB) showed the highest proportion of biome shifts (Figs. , S7). These relations were consistent for all RCPs (Fig. S7). The relation between the rate of biome shifts and the number of biomes in the biome map was positive but weak (Figs. , S7, Table S7), i.e., more biome types in the biome map imply a higher proportion of biome change. Yet, some biome maps with a lower number of biomes also showed high proportions of biome change.

The regression model showed, that the RCP scenario had a highly significant effect on the number of grid cells affected by biome change (t statistics <0.001, Table S8). The effects of the F31 biome map, GVM and the number of biomes in the map had lower effects, but still significant (t statistics between 0.01 and 0.05). The random forest regression model explained 76.4 % of the variance in the susceptibility to biome change (R2=0.764, RMSE=3.63%, Fig. , Table S9). Temperature anomalies had higher importance than precipitation anomalies. The interaction analysis revealed that temperature sensitivity was significantly higher in the cold zone (temperature <5 °C, p<0.001, Table S10). The partial dependency plots (Fig. S8) revealed a consistent non-linear increase in the susceptibility to biome change with rising temperature anomalies across all climatic zones. The susceptibility saturated at higher temperature increases for all climate zones (Fig. S8). Notably, the cold zone exhibited the highest absolute probability, compared to the temperate and warm zones. In the linear interaction model, the marginal effect of temperature anomalies on the susceptibility to biome change was 1.19 percentage points per °C for the cold zone. In contrast, the sensitivity in warm ecosystems was significantly lower, as indicated by a negative interaction term (β=-0.63,p<0.001), resulting in a marginal effect of only 0.55 percentage points per °C.

Our biome classification with random forests reproduced the biome map with high agreement (Fig. S9). The κ value ranged between 0.69 for random forests informed by CLM4.5 and 0.99 for models informed by , indicating substantial to almost perfect agreement (Tables S12 to S18). Biome-specific κ varied largely within models and for biomes between different models. For example, the values ranged between 0.40 (moderate agreement) and 0.91 (almost perfect agreement) for different biomes in ORCHIDEE, and between 0.1 (slight agreement) and 0.96 for “flooded grassland and savanna” between different GVMs (and 0.99 for the data). While the proportion of grid cells with data-model disagreement differed between the models, the regions overlapped substantially (Fig. S10). For example, disagreement occurred in areas bordering the central African rainforests, India or the west of North America where temperate conifer forest was not captured by some of the models (Fig. S10). Apart from the CARAIB model, κ values were higher for biomes with biome boundaries controlled by temperature, compared to biomes controlled by other factors (Table ). Yet, differences were not significant.

The five GVMs projected substantial biome transitions until the end of the century (Fig. ). The proportion of grid cells where all models projected biome shifts were 0.1 %, 0.3 % and 1.4 % of the land surface for RCP2.6, RCP6.0 and RCP8.5, while the proportion of grid cells not affected by biome change in any model were 51.6 %, 38.7 % and 47.9 % for the RCPs (grey area in Fig. ). For RCP6.0, the percent of grid cells affected by biome shifts ranged between 6.8 % for CLM4.5 and 37.1 % for CARAIB. The regions affected by biome change did not fully match, but we identified hotspots of potential biome shift. Regions where models consistently simulated biome transitions were scattered across different continents and biomes, for example along the “boreal forest” - “temperate broadleaved mixed forest” biome boundary or in the south of Brazil (Fig. S11). Further, most models projected transitions towards “tropical mixed broadleafed forest” in the regions bordering the central African rainforests, or transitions to “tropical grassland savanna and shrub” in India.

When considering all GVMs and RCPs, the most frequent biome transition was from “tropical and subtropical dry broadleaf forest” to “tropical and subtropical grassland savanna and shrubland” (43.0 %, 52.4 % and 52.9 % of grid cells affected by change for RCP2.6, RCP6.0 and RCP8.5, respectively, Fig. S12). Yet, coverage of “tropical and subtropical dry broadleaf forest” is low under current conditions. When considering only biomes with more than 5 % coverage under current conditions, the most frequent transitions were modeled between “tundra” and “boreal forest/taiga” (18.1 %, 27.8 % and 18.2 % affected, Fig. , Table S19). As hypothesized, the analyses showed that biome change was slightly more consistent, i.e., the coefficient of variation of areas affected by specific biome transitions was lower, for temperature-controlled biomes than for other biomes, but not significant (coefficient of variation 0.61 for temperature-driven biomes and 0.68 for others in RCP6.0, p-value 0.522 for t-test comparing coefficients of variation). Yet, the proportion of grid cells affected by change was smaller for temperature-driven biomes (Table ).

As expected, random forest models using GVM results and observation-based PFT maps showed high performance and agreement with the F31 biome maps (high κ value). The performance of RFs using PFT maps was higher than those using GVM results. This result is not surprising as PFT maps were derived from various remote sensing products , and they are therefore not fully independent from biome maps derived from similar remote sensing products . It is more remarkable that biome maps derived from GVM results showed very high agreement with observation-based biome maps (κ>0.8) for some combinations of GVM and the F31 biome maps. GVMs are process-based and bottom-up , and biome information is typically not directly used to parameterize such models. Yet, some models constrain current and future PFT cover fractions (e.g., CLM4.5) using observation-based PFT or biome data, instead of simulating PFT cover dynamically (e.g., LPJ-GUESS). However, LPJ-GUESS and other models use bioclimatic limits to constrain the distribution of PFTs, and these constraints also influence the possible distributions of biomes defined by those PFTs. Bioclimatic limits and responses to climate change differ between models. For example LPJ-GUESS, ORCHIDEE-DGVM, ORCHIDEE, and CLM4.5 mainly use temperature-based bioclimatic limits to constrain PFTs, whereas CARAIB adds explicit drougth tolerance constraints (Tables S2 and S3).

For GVMs, κ values were highest for the biome map. Classification in this map is based on vegetation, climate and other environmental factors, and re-evaluation of previous, regional-scale studies. The map shows a clear arrangement of biomes along environmental gradients and the relation to climatic drivers. High agreement with GVM results illustrates that biomes simulated by GVMs reflect the prevailing climatic conditions and bioclimatic limits that constrain the distribution of PFTs. Similarly, agreement between GVMs and the map, a map derived from species distributions and biogeographic zonation, was high. The map has often been used as reference map in biome-specific studies, for example, of the human footprint , climate change velocity , or biodiversity intactness . High agreement confirms that the biome map is suitable as a reference for such analyses.

Agreement between observation-based and modeled biomes was spatially heterogeneous. For example, in the map, agreement was lowest on the Tibetan plateau. also used elevation to define biomes, which is not used by the GVMs to simulate biomes. Other regions with low agreement were the sub-tropics surrounding the central African rainforests. It has been hypothesized, that in some of these regions, alternative biome states are possible , that is, depending on fire activity or herbivores, closed forests or open savanna states are possible . Fire has been included in many GVMs , including the models in our study, and shown to influence modeled tree cover and the carbon cycle . If fire enables alternative states and influences biome patterns in our results remains to be tested. Nevertheless, our study identified data-model mismatches for different GVMs and can be used to identify model limitations or to attribute uncertainties to fire, animals, humans or other factors. Doing such detailed analyses for the different models is, however, beyond the scope of our study.

Despite the large variation between different combinations of GVMs and F31 biome maps, κ values were overall higher than in previous studies. For instance, reports values between 0.2 and 0.79 in a comparison between different models and classification methods, and reports a maximum value of κ=0.52 for a classification using functional traits modeled by the aDGVM2. This can be attributed to the application of random forest classification that often shows high performance , while previous GVM studies typically applied expert-based biome classification approaches. Random forests maximize the agreement between modeled and observation-based biomes at the expense of clear and justifiable rules to separate between biomes. Further, we created a large ensemble of biome maps with κ values ranging between around 0.5 and >0.95.

All GVMs projected biome shifts until the end of the century, irrespective of the biome map used to inform the biome classification. Yet, the proportion of grid cells affected by biome change varied between different RCPs, GVMs and biome maps. Across all RCPs and GVMs, regions with high susceptibility of biome change were distributed across all continents, for example in southern Africa, the northern latitudes of Eurasia, or south of the Amazon rainforests. Biome stability or low susceptibility of biome change was consistently modeled in deserts and tropical rainforests globally.

Our results confirm some previous modeling results. In a global study, also identified deserts and tropical forests as biomes with lowest susceptibility of biome change and higher susceptibility in the northern latitudes. showed with LPJ-GUESS simulations that a large proportion of desert and rainforest have also been stable during the last 140 ka, showing their stability under past, current and future conditions (at least without direct human impacts). Yet, found that five out of seven climate-vegetation models project Amazon forest dieback at the local scale, that may, however, be compensated by increases in carbon elsewhere at larger scale. argued that under future warming, tropical forests may approach ecophysiological limits that inhibit further growth, particularly under RCP6.0 and RCP8.5. In combination with deforestation, warming may lead to tipping point behavior and transitions into degraded states . However, such effects may be compensated by adaptation and acclimation of temperature optima , or by high functional diversity buffing drought effects . Amazon forests may also experience more hot droughts and non-analogue climate under future conditions, enhancing the risk of forest dieback . Such large-scale die-back is not confirmed by our results and susceptibility to biome change was only low to moderate in the Eastern parts of the Amazon. The reasons for this discrepancy to other studies is not clear but likely related to different climate forcing data or to CO₂ fertilization that may overcompensate precipitation change in the GVMs. Such compensation effects have been observed, for example for Eucalyptus species grown in pots .

Biomes in the cold zone were generally more susceptible to warming than other zones. Our results for the biome map revealed that “temperate grassland savanna and shrubland” and “tundra” were highly susceptible to biome shifts in all RCPs. This result partly agrees with , who identified temperate mixed forests, boreal conifer forests, and tundra/alpine as most vulnerable to biome shifts. Similarly, showed with the LPJmL model that boreal forests are highly susceptible to dieback while temperate forests are relatively resilient and tropical forests are relatively stable.

While thermal acclimation and adaptation of temperature optima could mitigate some warming impacts , these processes are often not well represented in current GVMs. They often rely on fixed thermal response functions to model for example photosynthesis and respiration. Consequently, the lack of dynamic acclimation in models may lead to uncertainty in predicting the resilience of ecosystems under climate change . Ecopysiological processes and bioclimatic limits drive the performance of different PFTs at different temperatures. For instance, at low temperatures bioclimatic limits may constrain establishment of a specific PFT and unusual extreme cold can also kill trees, while at high temperatures respiration may reduce productivity. Climate change therefore has PFT-specific effects on the productivity, LAI, spatial coverage, and may thereby trigger biome transitions. In GVMs, the climatic niche of PFTs is generally preserved, and novel climates do not imply acclimation and adaptation. Whether novel and non-analogue biome states, characterized by novel combinations of PFT composition emerge in GVMs, remains to be tested. Uncertainties persist regarding the extent to which CO₂ fertilization offsets physiological stressors in emerging environmental niches and the models' ability to capture extreme heat and drought remains a key challenge for model development.

One reason for discrepancies between our and previous results is that predicted future large-scale dieback events are climate- and vegetation-model-specific e.g.,. For example, in LPJ-GUESS in this study the high-temperature stress function for boreal trees that LPJmL includes has been deleted more than ten years ago because the developers concluded that there is not enough observational data to parameterize this function. Therefore, LPJ-GUESS does not predict heat-related mortality and dieback in the boreal forest. Which model combination is most realistic is, at present, impossible to say, partly because it has hardly been tested to what extent the current GVMs can reproduce recent increases in drought- and heat-related tree mortality and extreme fires , such as the record fires in Canada in 2023 . Data-model comparisons for the recent sharp increase in drought-induced mortality of Norway spruce in Central Europe, in some regions reaching 50 % of the forest area, have shown that most GVMs fail to reproduce the observed large-scale forest die-back . Thus, the current GVMs might underestimate future risks of climate-induced forest dieback and associated biome shifts, but more studies will be necessary to corroborate such a conclusion. Given the limitations of the representation of tree mortality in vegetation models , improving the mortality in vegetation models will not only improve the representation of current ecosystem dynamics and biome boundaries, but also reduce uncertainties in projections of future climate change effects.

Our findings demonstrate that biome stability and biome shifts are primarily driven by temperature forcing, i.e., current temperature and temperature change. The higher sensitivity of biomes in the cold zone to warming suggests that these regions are approaching critical ecological thresholds. The saturation effect in the susceptibility to biome shifts beyond approximately 4 °C in the cold zone and 7 °C in the temperate and warm zone indicates a state of maximum climate-driven biome change that was reached at lower temperature change in the cold zone. These results underscore that vulnerability of biomes in the cold zone is not merely a function of higher warming under climate change but is caused by the higher intrinsic sensitivity of ecosystems to warming. In contrast to the tropics, acclimation may not necessarily buffer vulnerability to warming in cold and high-altitude biomes. Recent evidence suggests that alpine species possess a more limited capacity for thermal acclimation compared to other biomes .

For each GVM, the proportion of grid cells affected by future biome change decreased as data-model agreement (the κ value) with one of the 31 biome maps increased, i.e., higher agreement implies a lower proportion of grid cells affected by biome change. A similar relation was found in for the classification of aDGVM2 results into biomes using PFT cover fractions or patterns of functional traits modeled by the aDGVM2 . This result suggests that the rate of projected biome shifts may be overestimated when a biome classification with low data-model agreement is used. In contrast, biome classifications with high agreement may provide an estimate of the minimum rate of biome change, and potentially a more reliable trend of biome change. For RCP6.0, this lower limit is between 21.9 % and 28.9 % for the different GVMs.

Interestingly, assessing this relation across all GVMs included in our analysis showed the opposite, i.e., an increasing proportion of grid cells affected by biome change with higher data-model agreement. For better performing GVMs (i.e., higher κ) the random forest is more precisely defined and therefore more sensitive to LAI changes. Biome maps with low performance (i.e., poorer κ values) reflect weaker compatibility with the model PFTs, leading to spurious relationships and reduced sensitivity of the random forest models to LAI changes. CARAIB has the highest number of PFTs and the highest number of grid cells affected by biome change. A biome type may, in our classification, be represented by multiple PFTs, and moderate LAI changes of multiple PFTs may translate into a biome change. In models with a low number of PFTs, such as CLM4.5, a biome can be represented by a single PFT, and substantial changes in LAI may be required to trigger a biome change. In addition, PFT cover fractions in CLM4.5 remain static under future climates, such that biome change is constrained and only driven by LAI change.

Our results for the map support the hypothesis that biomes with well-described temperature limits, including for example tundra and boreal forest (Table S4), are better reproduced by the random forest models than biomes driven by other factors. Yet, differences between κ values of those groups of biomes were not statistically significant. Many GVMs include bioclimatic limits related to temperature to constrain the distribution of different PFTs e.g.,Table S3 for GVMs included in the study. These limits describe, for example, cold tolerance or chilling requirements of PFTs, and they are based on empirical evidences . Given the constraints on PFTs, bioclimatic limits also constrain the distribution of biomes. Yet, single PFTs do not match directly with biomes. Hence, it is not a priori clear, that GVMs using bioclimatic limits and ecosystem dynamics represented in GVMs will produce LAI and PFT patterns that represent the biomes included in different biome products. We therefore conclude that bioclimatic limits alone do not explain the good data-model agreement in our study. Thus, the border between forests and tropical and temperate grasslands is in most models not defined by bioclimatic limits and instead emerges from the competition between grasses and trees in interaction with fires.

Temperature-driven biomes showed lower proportions of grid cells affected by biome shifts than those controlled mainly by other factors. Yet, as expected, biome transitions were slightly more consistent for temperature-driven biomes (not statistically significant), i.e., the variability of possible future biome types for a given biome type under current conditions was lower between different models. This could be attributed to the representation of PFTs and bioclimatic limits in the models, but also to spatial patterns of changes in environmental conditions. Temperature is less variable among the climate models than other variables such as precipitation or aridity, and temperature changes are mainly directed poleward, in contrast to other variables . In the GVMs, temperature increases and elevated CO₂ in different RCP scenarios do not only modify ecophysiological processes, the carbon balance and competitive hierarchies of PFTs but temperature change may also imply that bioclimatic limits of PFTs are crossed (Tables S2, S3). Hence, climatic conditions may become unsuitable for the PFT, which implies poleward biome shifts, following the warming trends in the RCP scenarios. Changes in biomes that are not mainly driven by temperature indicate that effects of other variables such as change in rainfall regimes, CO₂ or fire activity have stronger effects on those biomes than change in temperature. Fire is included in many GVMs , including models used in our study. A recent analysis highlights the model's applicability for burned area attribution and further model comparisons focusing on fire are ongoing . Yet, we did not analyze fire effects on biome change.

The LPJ-GUESS version used in this study and some of the F31 biome maps e.g., only include natural vegetation, while other GVMs and biome maps also include cultivated areas. For consistency, we used only PFTs and biomes representing potential natural vegetation. Yet, direct human impacts such as deforestation of Amazon rainforests or intense livestock grazing in savannas have shaped the current biome distribution and can accelerate or inhibit biome change in the future . While land use, nitrogen deposition and fertilizer input were considered but fixed at year 2005 levels in our analysis, it could be repeated by including PFTs and biomes representing cultivated land and by using available ISIMIP results simulated with changes in land use, nitrogen deposition and fertilizer input according to HYDE3.2 for historic conditions and to MAgPIE simulations for SSP2 for future conditions . Considering only potential natural vegetation is, however, important for conservation and protection of the remaining natural vegetation.

We used LAI simulated by different GVMs for biome classification, for consistency with biome classification schemes used in previous modeling studies e.g., . LAI is directly modeled by GVM, observable and it has direct biophysical meaning. Hence, our approach provides a functional description of biomes, based on productivity and relative abundance of different PFTs. Selection of appropriate variables for biome classification is not only relevant for model results but also for observation-based biome maps. Depending on the variables used for classification and the number of biome types represented, the resulting biome maps can vary substantially . The F31 biome maps span six decades and various mapping approaches, from pre-satellite expert-based maps to modern products that integrate satellite remote sensing products. Direct comparisons of their agreement would require harmonization and re-classification . Our approach sidesteps such re-classification, by creating a random forest model for each F31 biome map.

While our random forest classification approach is reproducible and can be applied to any GVM, it has a major caveat: rules for classification are less transparent than in expert-based biome schemes tailored for specific GVMs and biome types. Yet, identifying thresholds to delineate between biomes using quantitative methods or expert knowledge has their own flaws. We therefore argue that large-scale model comparisons benefit from utilizing reproducible classification methods that can be applied to any set of PFTs included in GVMs. Expert-based classification schemes are more applicable for studies with single GVMs.

Vegetation change has been shown to lag behind change in environmental forcings e.g.,. Hence, vegetation and biome patterns are committed to further change to reach an equilibrium state with climate, even if the climate system stabilizes. These lags can be explained by various processes including dispersal limitation, delayed responses of ecological processes such as ecophysiology, establishment and mortality, succession, and disturbances such as drought or fire. Dispersal has been integrated into GVMs e.g.,. Yet, consideration at global scale is challenging as dispersal is strongly influenced by small-scale processes and heterogeneity that cannot be represented at the 0.5° resolution of the GVM simulations . In our analysis, we ignored potential biome shifts between the publication of the different observation-based biome maps and our reference period for the analyses, and respective temporal mismatches or lag effects. Nonetheless, random forests based on GVM results reproduced broad biogeographic patters encoded in the different observation-based biome maps.