To assess SM-induced changes in the land carbon sink, ESM output data of LFMIP are used, which is part of the Land Surface, Snow, and Soil Moisture Model Intercomparison Project LS3MIP, under CMIP6. LFMIP follows the GLACE-CMIP5 blueprint and is designed to diagnose changes in land–atmosphere coupling related to SM.

LFMIP consists of three experiments, a reference run (CTL) based on the historical and the Shared Socioeconomic Pathway 5-8.5 (SSP5-8.5) scenarios of CMIP6 and two experiments where SM is prescribed as (i) the mean seasonal cycle of 1980–2014 (pdLC) and (ii) the 30-year running mean (rmLC) of CTL (see Supplement Fig. S1 for an illustration of the experiments). LFMIP outputs are available from the four ESMs of CMIP6, CESM2, IPSL-CM6A-LR, MPI-ESM1-2-LR, and CMCC-ESM2 (for detailed information see Supplement Table S1). LFMIP covers the period 1981–2099 at monthly resolution.

It is important to note that for the experiment simulations pdLC and rmLC, the hydrological cycle is not closed because by prescribing SM externally, hydrological feedbacks from the atmosphere to SM are suppressed. Consequently, water can artificially be added or lost. However, as our focus lies on the SM-induced impact on the carbon cycle and not on the hydrological cycle itself, all relevant processes, including changes in surface energy partitioning, near-surface air temperature, VPD, and physiological controls on photosynthesis remain physically consistent. Therefore, differences between experiment simulations can be robustly attributed to SM-induced effects on NBP and GPP despite the lack of water budget closure.

The SM-induced changes in NBP derived from LFMIP are compared with those of the previous generation, GLACE-CMIP5 (Table ). For this purpose, we partially reproduce and build on the analysis of to compare SM-induced changes in NBP between model generations for the period 1981–2099. Available for the analysis of GLACE-CMIP5 are outputs from four models, CESM, GFDL, IPSL, and MPI-ESM (see Table 1 for detailed information). Among the available ESMs for each model generation, three of four model are from the same modelling group. The GLACE-CMIP5 experiment follows the same protocol as LFMIP for the previous model generation, CMIP5. To compare recent and projected future conditions across models, our analysis focuses on a baseline period (1981–2010) and an end-of-century future period (2070–2099). We note that the period for the SM prescription in GLACE-CMIP5 (1971–2000) is roughly one decade earlier than in LFMIP. This impedes the direct comparison for the impact of SM on NBP across model generations for the baseline period, but our focus lies on the evolution of SM-induced impacts on NBP by the end of the century, which are not affected.

To assess if the available four models of LFMIP and GLACE-CMIP5, respectively, are representative of their respective CMIP generation, larger sets of CMIP6 and CMIP5 models are used. Ten ESMs of CMIP5 (i.e., BNU-ESM, CESM1-BGC, CanESM2, CanESM2, GFDL-ESM2G, HadGEM2-CC, HadGEM2-CC, IPSL-CM5A-LR, MPI-ESM-LR, and NorESM1-M) and nine ESMs of CMIP6 (i.e., ACCESS-ESM1-5, CESM2, CMCC-ESM2, EC-Earth3-Veg-LR, GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-LR, MRI-ESM2-0, NorESM2-LM) were used for this comparison, selecting one ensemble member from each modelling group.

ESMs are physically based, prognostic models that include a land component with an interactive carbon cycle described by numerical representations of real−world biogeochemical processes. GPP is simulated as leaf level photosynthesis and constrained by temperature, radiation, atmospheric CO₂ concentration, water and nutrient availability. ESMs also represent soil biogeochemistry, including litter and soil organic matter decomposition that drive microbial respiration. NBP is provided as the net land–atmosphere carbon exchange (i.e., NBP = GPP - ecosystem respiration - disturbances).

We use ESM output for NBP to assess the land carbon sink, GPP to assess land carbon uptake, and total column soil moisture to quantify changes in projected SM. Since ESMs account for different numbers of soil layers and also vary in soil layer depth, output data for SM varies substantially in magnitude. Therefore, we standardise SM output to z-score values, using 1z=X-μrefσref where each data point X is standardised by the mean of the reference period μref and its standard deviation σref. As reference period we use the baseline period (1981–2010) to emphasize changes in SM compared to the baseline period from which SM is prescribed in the experiments.

For the contribution analysis of direct and indirect SM effects (see Sect. ) we further use LFMIP model output at monthly resolution for the variables near-surface air temperature (T, in °C), surface downwelling short-wave radiation (R, in W m⁻²) and vapour pressure deficit (VPD, in kPa). VPD is calculated as the difference between the saturated water vapour pressure (es, in kPa) and the actual water vapour pressure (ea, in kPa), where es is determined by T using the Clausius–Clapeyron relation, such that es=0.6108×exp⁡17.27TT+237.3 and ea is calculated from relative humidity (RH, in %) as ea=RH100×es.

For spatially resolved comparisons across models, the data is regridded to 1° resolution using bilinear interpolation.

To isolate the effects of SM on NBP (and similarly, on GPP), a method commonly used in experiments following the GLACE blueprint is employed (see . Changes in NBP (ΔNBPCTL) can be described as 2ΔNBPCTL=ΔNBPSMtrend+ΔNBPSMvar+ΔNBPother where ΔNBPSMtrend is the effect of changes in mean SM conditions on NBP and ΔNBPSMvar the effect of SM variability on NBP. ΔNBPother summarises changes in NBP due to the fertilisation effect of CO₂, temperature changes, and other influencing factors. The experiments of LFMIP allow isolating the effect of SM trend and variability, where 3ΔNBPSMtrend=ΔNBPrmLC-pdLC and 4ΔNBPSMvar=ΔNBPCTL-rmLC, as well as the combined effects of SM expressed as 5ΔNBPSMall=ΔNBPSMtrend+ΔNBPSMvar=ΔNBPCTL-pdLC.

To assess the impact of direct and indirect SM effects, we focus on GPP which is more strongly influenced by SM changes than respiration, particularly in regions that strongly contribute to global NBP (see Supplement Figs. S2 and S5). SM-induced changes in GPP (ΔGPP) can occur as direct effect through changes in water availability for photosynthesis and as indirect SM effect via SM–atmosphere coupling. To assess the contribution of direct and indirect effects, we conduct a contribution analysis (see Sect. 3.2.1). Following the approach of , a multiple linear regression is performed to capture the local direct and indirect effects of SM on GPP for monthly time steps on grid cell level, using 6ΔGPP*=βSMΔSM+βTΔT+βVPDΔVPD+βRΔR, where ΔGPP* describes the estimated SM-induced change in GPP due to ΔSM, which represents the change in SM; ΔT denotes the SM-induced change in near-surface air temperature, ΔVPD that in vapour pressure deficit, and ΔR that in downward solar radiation, respectively, and β are the corresponding regression coefficients. The regression is performed over 30 years and the resulting impacts of the individual drivers SM, T, VPD, and R are referred to as ΔGPPSM*, ΔGPPT*, ΔGPPVPD*, and ΔGPPR*, respectively. Following the reasoning of , the impact of indirect SM effects via T and VPD (ΔGPPT* and ΔGPPVPD*) are reported as their sum (ΔGPPT&VPD*), since the calculation of VPD depends on T, limiting the ability to disentangle their impact on GPP. Their individual effects are reported in Supplement Figs. S12 and S13 for completeness.

We note that for this part of the analysis we focus only on the total SM effect (and not on the individual effects of SM trend and variability), thus, ΔGPP is equivalent to GPP_CTL - GPP_pdLC and consequently the regression estimates also refer to the total SM effect on GPP. For better readability we omitted the subscript “SMall” for this part of the analysis.

To assess whether differences in projected direct and indirect SM effects or in the sensitivity of GPP to those SM effects dominate intermodel spread, we apply a variance decomposition based on an analysis of variance (ANOVA) framework to the SM effects derived by regression, i.e., ΔGPPSM*, ΔGPPT*, and ΔGPPVPD* (see Sect. ). We neglect ΔGPPR* due to its small impact on global ΔGPP* (see Sect. ). We construct a full factorial ensemble of ΔGPP* by systematically combining the projected SM effects and sensitivities of GPP across LFMIP models. At each grid cell, the SM-induced change in GPP for a given combination of factor levels is expressed as 7ΔGPPijklmn*=βSMiΔSMj+βTkΔTl+βVPDmΔVPDn+εijklmn, where ΔSM, ΔT, and ΔVPD denote the direct and indirect SM effects and βSM, βT, and βVPD are the sensitivities of GPP to the respective SM effect. The indices i,j,k,l,m,n denote the levels for each of the six factors (i.e., the individual ESMs) from which the SM effects and sensitivities are drawn. ε represents the residual change in GPP which is not captured by the linear contributions of direct and indirect SM effects, including non-linear, non-local, lagged, and interaction effects. The ANOVA-based variance decomposition then attributes the variance in ΔGPP* across factor combinations to (i) the sensitivity of GPP to SM, T, and VPD and (ii) the projected changes in SM, T, and VPD. This decomposition allows us to quantify the relative contributions of differences in projected SM effects and in the associated sensitivities of GPP to the intermodel spread of SM-induced impacts on GPP.

We assess the impact of SM on NBP for the latest CMIP generation using the available LFMIP models, and compare the results with those from GLACE-CMIP5 (Fig. , Supplement Figs. S3, and S4). The LFMIP model mean projects the global land carbon sink (global NBP_CTL) to increase from 1.07 ± 0.11 PgC yr⁻¹ (model mean ± standard deviation across models) during the baseline period (1981–2010) to 2.91 ± 1.38 PgC yr⁻¹ by the end of the century (2070–2099, Fig. a and c). The GLACE-CMIP5 mean for global NBP_CTL is of similar magnitude as that of LFMIP for the baseline period (1.23 ± 1.38 PgC yr⁻¹). However, GLACE-CMIP5 projects a decline in global NBP_CTL after 2070, resulting in only a small increase in the land carbon sink by the end of the century (1.51 ± 1.69 PgC yr⁻¹ (Fig. b and c)), whereas LFMIP projects the global NBP_CTL to almost triple. Thus, the LFMIP mean projects a stronger land carbon sink by the end of the century compared to the GLACE-CMIP5 mean. However, the significantly smaller global NBP_CTL in GLACE-CMIP5 is influenced by the CESM model, which projects NBP to be a consistent net carbon source over the 21st century (Supplement Fig. S3a).

In both model generations, projected SM changes negatively affect global NBP_CTL (Fig. ). For the LFMIP mean, the projected negative effect of SM on NBP is overall less severe compared to GLACE-CMIP5 (Fig. a and b). GLACE-CMIP5 models project the global land carbon sink to be reduced by about twice its absolute magnitude during the baseline period (2.44 ± 1.77 PgC yr⁻¹), as previously shown by (Fig. b). This is primarily due to the negative impact of SM variability (-2.21 ± 1.67 PgC yr⁻¹). While the impact of the SM trend leads to a growing reduction over the 21st century (to -1.14 ± 1.09 PgC yr⁻¹ by the end of the century), the negative effect of SM variability is projected to decline (to -1.26 ± 0.41 PgC yr⁻¹ by the end of the century, Fig. c). The total impact of SM on NBP at the end of the century reduces the global land carbon sink to half (i.e., the land carbon sink would be twice as strong without the negative impact of SM). In LFMIP, the projected negative effect of SM on NBP is overall less severe (Fig. a). In the baseline period, SM changes reduce the global land carbon sink by -0.69 ± 0.15 PgC yr⁻¹. In the future period, the negative impact of SM changes increases in absolute magnitude to -1.22 ± 1.04 PgC yr⁻¹ (Fig. c). The effect of SM variability remains relatively unchanged throughout the century, reducing NBP by -0.66 ± 0.22 PgC yr⁻¹ in the baseline period and -0.82 ± 0.80 PgC yr⁻¹ in the future period. The impact of SM trend becomes increasingly negative, resulting in a reduction of -0.40 ± 0.27 PgC yr⁻¹ by the end of the century. Nevertheless, in relative terms the negative impact of SM reduces global NBP_CTL by 64 % for the baseline period and only by 42 % during the future period, because global NBP_CTL of LFMIP is projected to increase over time.

Assessing spatial patterns of the effect of SM on NBP shows a reduction of NBP due to SM in most areas for both model generations, mainly due to the effect of SM variability (Fig. , Supplement Fig. S6). The reduction is strongest in tropical regions of South America and in the mid-latitudes of the northern hemisphere, where NBP_CTL is generally high (Fig. , Supplement Fig. S5). This SM-induced reduction in large parts of the globe is already apparent during the baseline period due to the negative impact of SM variability (Supplement Figs. S8 and S9). While LFMIP projects the negative effect of SM on NBP to intensify further over the century, especially in tropical regions and mid-latitudes of the northern hemisphere (Fig. a, Supplement Fig. S9), this is not the case in GLACE-CMIP5, where several regions show a substantial reduction in the negative effect of SM, especially in the mid and high latitudes of the northern hemisphere (Fig. b, Supplement Fig. S8). However, in both CMIP generations, models show high disagreement on the sign of change of the SM-induced impact on NBP for the future period relative to the baseline period in several regions, including regions where models agree on the projected change in SM (Fig. ). For both CMIP generations, intermodel differences in SM-induced changes in NBP are mainly located in tropical regions and the northern mid-latitudes.

Comparing the SM evolution in GLACE-CMIP5 and LFMIP reveals substantial differences in the magnitude and spatial extent of the projected SM drying (Fig. ). The GLACE-CMIP5 mean shows widespread SM drying across the globe (Fig. a.1), which is less pronounced in LFMIP (Fig. b.1). The GLACE-CMIP5 models IPSL and GFDL (Fig. a.3 and a.5) project severe SM drying in several areas, especially in the northern mid-latitudes by the end of the century, resulting in a strong negative impact of SM on NBP. The CMIP6 version of IPSL included in LFMIP, IPSL-CM6A-LR (Fig. b.3), shows less SM drying in most areas than IPSL and even a reversed trend (i.e., SM wetting) in several regions, including vast areas of Central and Southeast Asia, Northern Europe, Central Africa, and North America. Similarly, MPI-ESM1-2-LR (LFMIP, Fig. b.4) shows less pronounced SM drying than MPI-ESM (GLACE-CMIP5, Fig. a.4) and a reversed SM trend in large parts of Central and Southern Africa, as well as in Western and Southeast Asia, leading to a less negative SM-induced impact on NBP. CESM2 (LFMIP, Fig. b.2) and CESM (GLACE-CMIP5, Fig. a.2) generally show less pronounced differences in their SM projections.

We assess the representativeness of the four available models in GLACE-CMIP5 and LFMIP for the respective CMIP generation by comparing them with multi-model means (MMMs) of CMIP5 and CMIP6, respectively. The comparison was performed for NBP and SM for each latitude zone (Fig. ). LFMIP and GLACE-CMIP5 have a similar mean and spread (95 % confidence interval) as the respective full ensemble in most latitude zones. However, GLACE-CMIP5 shows stronger SM drying in the mid-latitudes of the northern hemisphere than the CMIP5 MMM, primarily due to GFDL and IPSL (Fig. b.1, Supplement Fig. S10b.2). As the northern mid-latitudes are the main contributors to global NBP (accounting for 41 % (45 %) of global NBP in LFMIP (CMIP6 MMM) and 70 % (85 %) in GLACE-CMIP5 (CMIP5 MMM)), this must be considered when interpreting the strong negative impact of SM on NBP projected by these models in GLACE-CMIP5.

With focus on LFMIP, we further analyse the origins of SM-induced changes in land carbon uptake by (i) conducting a contribution analysis to assess the impact of direct and indirect SM effects on GPP (Sect. 3.2.1) and (ii) investigating causes of intermodel differences in SM-induced changes in GPP from direct and indirect SM effects and their respective sensitivities (Sect. 3.2.2). To assess the impact of SM on land carbon uptake under future climate change, we focus on the future period (2070–2099) when impacts of both SM variability and SM trend come into play. The results for GLACE-CMIP5 are displayed in Supplement Figs. S15 to S17.

Our results show that projected SM changes negatively impact global NBP according to the LFMIP mean, which is consistent with previous findings based on GLACE-CMIP5 . The negative effect is dominated by SM variability, highlighting the strong negative impact of dry SM extremes on NBP. This finding aligns with observational evidence of the severe negative impact of extreme drought and heat events that lead to large carbon losses of terrestrial ecosystems . However, the negative SM effect on NBP is much stronger in the GLACE-CMIP5 mean, which is substantially influenced by two models, IPSL and GFDL, which also project particularly severe SM drying. Comparing SM projections of GLACE-CMIP5 to the full CMIP5 ensemble shows that this drying signal is not representative of the model generation, as both models exhibit a strong drying in the mid-latitudes of the Northern Hemisphere not reflected in the CMIP5 MMM. The NBP in the northern mid-latitudes makes up for about 85 % of the total global NBP for the CMIP5 MMM. Consequently, the stronger SM-induced reduction of NBP in GLACE-CMIP5 may be partly due to the model subset's bias toward stronger drying in these latitudes, suggesting that differences in the projected SM impact on NBP between GLACE-CMIP5 and LFMIP might thus be partly attributable to model selection rather than a systematic differences between model generations.

Nevertheless, this does not imply that the overall conclusions of prior studies are overstated, because structural model limitations may lead to an overall underestimation of the severity of SM-induced reductions of NBP in ESMs of both generations. Previous studies highlighted that SM drought can strongly reduce carbon uptake, often followed by declining tree growth rates and increased tree mortality and potentially leading to lasting reductions in carbon uptake capacities of affected ecosystems . This is often not well captured in ESMs because limitations in the representation of fundamental plant hydraulics can lead to shorter and weaker simulated drought impacts on GPP than observations indicate . For example, some CMIP5 models (including CESM and GFDL, which participated in the GLACE-CMIP5 experiment) do not sufficiently capture the lagged effects of SM on NBP compared to observational data, although GFDL was shown to capture these effects best of all the CMIP5 models assessed . In addition to the projected strong SM drying, this may further explain the strong negative impact of SM on NBP our study reveals for GFDL, but further research would be needed to confirm this. In addition, declining moisture availability has also been linked to increased risk of fire , but ESMs have been found to misrepresent the observed magnitude of carbon losses during fire events . With compound fire weather and drought events increasing under climate change , this may lead to carbon losses not captured by ESMs.

As the occurrence of negative SM extremes is projected to increase under climate change , the inability of models to accurately capture these processes could lead to a growing underestimation of the severity of the negative SM-induced impacts on the future land carbon sink. Furthermore, demonstrate that the increased water-carbon coupling observed in the tropics (which implies increased carbon loss with declining water availability) is not well captured in state-of-the-art ESMs. Since ecosystems are projected to experience widespread shifts from energy to water limitation throughout the 21st century , this may further contribute to an underestimation of the simulated carbon loss under water stress.

Experiments with prescribed SM, liked those performed under GLACE-CMIP5 and LFMIP, offer a unique opportunity to isolate and assess the impact of SM changes on the evolution of the land carbon sink. Such information cannot be derived from observations because SM changes co-vary with other factors influencing the land carbon sink, thus isolating SM effects would require large-scale controls to fully capture SM–atmosphere interactions. However, we note that the validity of our results is limited by the small number of ESMs participating in LFMIP (and GLACE-CMIP5).

Conducting GLACE-like experiments with a larger set of models would enhance the robustness and validity of the results and provide further insights into the effects of SM on the land carbon sink and associated uncertainties under future climate change. We therefore recommend continuing GLACE-like experiments in future CMIP generations with more models to generate a larger ensemble. Furthermore, it has been shown that SM evolutions differ substantially across ESMs, due to differences in model process implementation , which complicates intermodel comparison. We therefore propose to extend the experimental setup by an experiment with identically prescribed SM evolution (e.g., prescribed trend and variability based on the CMIP MMM). This would enable isolating the effects of identical SM trend and variability on NBP (while the existing GLACE-like experimental setups allow isolating only the model-specific effect of SM trend and variability), by explicitly disentangling the contributions of differences in model sensitivity of GPP to SM from intermodel deviations in the simulated SM itself. Since our results from the ANOVA-based variance decomposition indicate that differences in the sensitivity of GPP to SM are the main origin of intermodel uncertainty, this approach would allow for a more direct comparison of the representation of water-carbon coupling across ESMs.

Using benchmarking frameworks such as ILAMB helps to assess model performance by systematically comparing model output to observational datasets. Overall, CMIP6 models show improved skill relative to CMIP5 in representing key components of the carbon and hydrological cycles, indicating that model accuracy has been improved . However, water-carbon coupling emerges from tightly interacting processes across soil, vegetation, and atmosphere. As land surface models have grown in structural complexity , evaluating individual variables in isolation may not reflect how well models capture underlying process interactions. Indeed, even when models perform well for SM and GPP individually according to the ILAMB broad scores, the coupling between SM and GPP may still be misrepresented . Expanding benchmarking approaches to explicitly target soil moisture-carbon interactions, dryness stress responses, and regime shifts between energy- and moisture-limited conditions would provide valuable constraints to assess model performance regarding SM-induced effects on NBP.

However, benchmarking simulated SM and its impact on the land carbon sink against past observations can only provide limited insights about satisfactory model performance under climate change in the future, particularly since the effects strongly depend on the forcing scenario and human perturbations of the water and carbon cycles . Assessing and improving the representation of water-carbon interactions under emerging climate states therefore remains a key challenge for the next generation of ESMs.