Net release of CO2 from thawing permafrost soil carbon predicted to occur earlier in this century Authors: Yi Xi, Philippe Ciais, Dan Zhu, Chunjing Qiu, Yuan Zhang, Shushi Peng, Simon P. K. Bowring, Daniel S. Goll, Pierre Friedlingstein, Gustaf Hugelius Published: 2026 ABSTRACT Accelerating permafrost thaw may release vast deep (>3 meters) frozen soil carbon as carbon dioxide (CO2), but this magnitude remains uncertain because current Earth system models (ESMs) lack deep carbon processes. Using an updated ORCHIDEE-MICT model simulating Pleistocene Yedoma formation and Holocene peatland development, we project northern (>30°N) carbon responses under climate change. Compared to the original model, including these deep carbon pools improves agreement with observations and reduces net CO2 uptake by 47 to 74 petagrams of carbon from 1900 to 2100 across three future scenarios because of deep carbon decomposition with accelerated active-layer deepening. Under high-emission pathways, the northern soil carbon balance shifts from a sink to a source of 32 petagrams of carbon, advancing the reversal reported in earlier studies into the 21st century. Consistent with field data, our model shows that colder soils retain more labile carbon—contrary to assumptions in many Coupled Model Intercomparison Project (CMIP) models—helping explain their persistent sink bias. Our results highlight the need to represent both the quantity and quality of permafrost carbon in ESMs. INTRODUCTION The northern permafrost region holds nearly one-third (1460 to 1600 Pg C) of the world’s soil organic carbon (SOC) (1–3) and is warming two to four times faster than the global average (4, 5). The response of this vast SOC pool to ongoing climate change remains a critical concern (6–9). Recent compilations of year-round carbon dioxide (CO2) flux measurements across the permafrost region reveal that previously undermonitored nonsummer CO2 emissions have increased substantially over the past decade, causing more than 30% of permafrost ecosystems, primarily tundra, to function as net CO2 sources (10–12). Intensifying disturbances such as abrupt thaw (13, 14) and wildfires (15), which are not accounted for in CO2 flux datasets, are expected to further enhance permafrost CO2 emissions (12, 16, 17). Because of inadequate representations of permafrost carbon processes, current Earth system models (ESMs) from the Coupled Model Intercomparison Project Phase 6 (CMIP6) fail to capture the contemporary CO2 source regions in the permafrost area (12). Instead, they project continuous soil carbon accumulation over the coming century, driven by the dominant response of plant productivity to warming (18). Before incorporating additional processes (e.g., wildfires and abrupt thaw) into ESMs to improve the prediction of northern soil carbon dynamics, it is equally important to refine existing modeling of permafrost carbon processes (19). Most CMIP6 models still lack critical permafrost-specific features in their soil carbon modules, such as vertically resolved decomposition along the SOC profile, freezing-induced decomposition inhibition, cryoturbation-driven SOC mixing, and SOC thermal insulation effects (7, 20). These omissions not only lead to an underestimation of permafrost extent and high-latitude carbon stocks against observation-based datasets in most ESMs but also hinder our understanding of how active layer thickening may influence permafrost carbon decomposition in the future (20, 21). In particular, deep carbon deposits, reaching ~10 m in depth in peatlands and ~20 m in Yedoma deposits (ice-rich, organic-rich permafrost), remain missing in CMIP6 models despite their substantial SOC storage of 415 (±150) to 547 (±74) Pg C for northern peatlands (6, 22) and 210 (±70) to 456 (±45) Pg C for Yedoma (23, 24). These deep frozen carbon reservoirs [all Yedoma (25) and 63% of northern peat (6) are underlain by permafrost (26)] also contain a high proportion of labile soil carbon, which is currently preserved from decomposition but may be rapidly mobilized upon thawing (6, 27, 28). While projections of active layer thickening under future climate change are termed as “highly confident” (18), the extent to which the increasing exposure of deep, labile carbon to higher temperatures will accelerate permafrost carbon loss remains uncertain. To better understand the interplay between permafrost thaw and deep carbon decomposition, we have recently proposed a new, more realistic SOC accumulation simulation protocol before the Industrial Era and human-induced carbon cycle disturbances. This protocol accounts for (i) sedimentation processes forming Yedoma deposits since the last glacial period and (ii) the transient development of northern peatlands during the Holocene based on peat age reconstructions. It was implemented using ORCHIDEE-MICT (ORganizing Carbon and Hydrology in Dynamic EcosystEms–aMeliorated Interactions between Carbon and Temperature), a land surface model specifically designed for high-latitude processes that includes critical permafrost-specific features (29, 30). Further details on the implementation of this protocol and its evaluation against observations can be found in the work of Xi et al. (31). We conducted simulations of SOC accumulation up to a quasiequilibrium state in the preindustrial era across the northern lands (>30°N) using both the new protocol and the standard (or “old”) protocol commonly used in current ESMs (fig. S1). The old protocol assumes that SOC reaches a preindustrial equilibrium without contributions from Yedoma carbon and that peatland SOC accumulation begins uniformly 13,500 years ago. By contrast, the new protocol simulates transient Yedoma carbon accumulation beginning 21,000 years ago and peat carbon accumulation with spatially variable initiation times. These spin-up simulations were followed by historical runs for the period of 1900 to 2014 and future projections for the period of 2015 to 2100 under three shared socioeconomic pathway scenarios (SSPs) (see Materials and Methods and table S1). By comparing the outputs of the new protocol simulations (referred to as SIMnew hereafter) with those from the old protocol (SIMold), we find that incorporating deep carbon substantially reduces the projected net sink of northern soil carbon between 1900 and 2100 and even reverses the projected soil carbon balance from a net sink to a net source under the high-emission scenarios. RESULTS Change in preindustrial total organic carbon with deep carbon processes Figure 1A shows the spatial distribution of total organic carbon (TOC; comprising litter organic carbon and SOC) across the northern region (>30°N), accounting for Yedoma and peatlands deposits under the new protocol. The high-latitude permafrost, Yedoma, and peatland regions—the spatial ranges of which often overlap—constitute the primary simulated soil carbon hotspots, with substantially higher TOC values (>40 kg C m−2, expressed per unit area across the full 38-m soil column in the model) compared to soils in lower latitudes. Our TOC estimate of 2028 Pg C is 226 Pg C higher than in the old protocol. Almost all (246 Pg C) of the additional TOC is concentrated in the permafrost region (Fig. 1B). When splitting additional TOC by source plant functional type (PFT), the Yedoma PFT and conventional PFTs (15 non-Yedoma and nonpeat PFTs; see Materials and Methods) contribute 70% (157 Pg C) and 30% (68 Pg C) of the additional TOC, respectively, while TOC in the peat PFT remains unchanged between the two protocols (Fig. 1C). Across soil depths, 64% (145 Pg C) of the additional TOC is stored within the shallow soil layers (0 to 1 m and 1 to 2 m), while 39% (88 Pg C) is found below 3 m (Fig. 1D). With respect to model carbon pools (see Materials and Methods for details of the soil carbon module), the additional TOC brought by the representation of Yedoma and peat is primarily under the form of fast-turnover pools—the “active” and “slow” pools—which have turnover times at 5°C without moisture limitation (τpotential_5°C) of 0.84 and 31 years, respectively. The TOC of these two pools is 47 Pg C and 227 Pg C higher under the new protocol, while the “passive” SOC pool (τpotential_5°C = 1363 years) is 52 Pg C lower (Fig. 1E). Change in preindustrial TOC resulting from the representation of deep carbon evolution across the northern region (>30°N). (A) Spatial pattern of TOC in the full 38-m soil column simulated using the new protocol, with the distribution of permafrost from Brown (26), Yedoma from Strauss et al. (25), and peatland from Hugelius et al. (6). (B to E) Comparison of simulated TOC under the old and new protocols by region (B), by PFT (C), by soil layer depth (D), and by model carbon pool (E). Three key mechanisms drive the preindustrial TOC change resulting from the more realistic representation of deep carbon accumulation in the new protocol (31). First, the inclusion of Yedoma deposits with their high SOC density (28 kg C m−3 excluding excess ice, expressed per unit soil volume) and thick depth (~20 m) substantially increases the TOC and the stocks of deep carbon below 3 m across the northern region. Second, the new protocol incorporates the insulation effect of the progressively increasing SOC on soil thermal processes during the long-term accumulation of SOC. This effect cools soil temperature, thus promoting SOC accumulation, particularly in high-latitude cold zones (section S1). Last, the new protocol accounts for realistic transients of peatland formation during the Holocene, which shortens the mean peat carbon accumulation duration by 4800 years compared to assuming a uniform spin-up period of 13,500 years for peat. This results in shallower peat carbon depth and reduced formation of long-turnover-time passive SOC in northern peatlands. Regionally, the contribution of the three mechanisms to the TOC change differs (fig. S2). In northern Siberia, the widespread distribution of Yedoma drives a substantially higher TOC (126 Pg C), accounting for 74% of the additional TOC in Asia under the new protocol (fig. S2, A to E). In northern Europe, where Yedoma is absent and the major peatland complexes are older than 10,000 years (fig. S3), the SOC thermal insulation effect enhances peat carbon storage by 20 Pg C (fig. S2, F to J). In North America, while Yedoma in Alaska and high-latitude conventional PFTs contribute to TOC increases of 29 Pg C (4%) and 41 Pg C (6%), respectively, the TOC reduction (−46 Pg C) in the peat PFT due to the shortening peat carbon accumulation in young peatlands (fig. S3) limits net TOC increases to just 4% for this continent under the new protocol (fig. S2, K to O). In addition, peat carbon below 3 m in depth and passive peat SOC decline by 46 Pg C (39%) and 16 Pg C (50%), respectively. Despite these distinct driving mechanisms, however, all three subregions show net increased TOC in shallow soil layers and labile carbon pools in the ORCHIDEE-MICT simulation using the new protocol, suggesting a potential for enhanced soil carbon decomposition across the northern region after the preindustrial era. Change in simulated historical TOC On the basis of the simulated preindustrial SOC by ORCHIDEE-MICT using the old and new protocols, we performed historical simulations for 1900 to 2014 and future projections for 2015 to 2100 across the northern region (>30°N) using bias-corrected atmospheric climate data from the ESM IPSL-CM6A-LR within the Inter-Sectoral Impact Model Intercomparison Project 3b (table S1 and Materials and Methods) (32). Between 1900 and 2014, the IPSL climate model simulates 2.3°C warming across the northern region (fig. S4A), which may overestimate historical warming by ~35%, as inferred from a comparison with the fifth generation ECMWF reanalysis dataset (ERA5) (33) for their overlapping period (1979 to 2014) (fig. S5A). In response to this warming, both SIMnew and SIMold show a net TOC gain over this period, although the TOC gain in SIMnew (32 Pg C) is 33% lower than in SIMold (48 Pg C) (Fig. 2A). This lower historical SOC gain in SIMnew is primarily from conventional PFTs (51%, −8 Pg C), then Yedoma PFT (31%, −5 Pg C), and peat PFT (18%, −3 Pg C). Regionally, 89% of this lower historical TOC gain in SIMnew occurs within the permafrost region (Fig. 2, E and F, and fig. S6), where nearly all of the simulated TOC difference between the two protocols is found (Fig. 1B). Temporal variations in TOC across the northern region (>30°N) from 1900 to 2100 (relative to 1900). (A to F) Temporal variations in TOC for all PFTs (A), conventional PFTs (B), peat PFT (C), Yedoma PFT (D), permafrost region defined by Brown (26) (E), and nonpermafrost region (F) from simulations using the old protocol (SIMold) and the new protocol (SIMnew). Three future scenarios are included: SSP1-2.6, SSP3-7.0, and SSP5-8.5. Evaluation of simulated historical terrestrial ecosystem CO2 fluxes Before analyzing the change in future SOC projections with our revised representation of deep carbon accumulation, we first evaluated simulated historical terrestrial ecosystem CO2 fluxes against a machine learning (ML)–based upscaling product. This product, to our knowledge, compiled the most extensive in situ CO2 flux observations across the Arctic-boreal zone [ABZ; including both boreal forests and tundra ecosystems as in the work of Dinerstein et al. (34); see fig. S5, G, I, and K] (12). Although with a 30% overestimation, the model generally captures the observation-based, multiyear mean annual and seasonal gross primary productivity (GPP), ecosystem respiration (Reco), and net ecosystem productivity (NEE) during the common period of 2001 to 2014 across the ABZ and its two subregions (Fig. 3, A to C, and figs. S7 and S8). Specifically, both simulations and the ML-upscaling flux model indicate that the ABZ functioned as a net annual carbon sink (−0.8 to −0.6 Pg C year−1) from 2001 to 2014, with the summer season (June to August) acting as a carbon sink (−2.6 to −2.2 Pg C year−1) and the nonsummer season (September to May) as a carbon source (1.6 to 1.8 Pg C year−1). The boreal ecosystem, which covers a larger portion of the ABZ (61%) and has higher vegetation productivity than the tundra, dominates the region’s carbon budget in both simulations and the ML upscaling (Fig. 3C). For the tundra ecosystem, the ML-based product estimates a small carbon source (0.07 Pg C year−1) during 2001 to 2014, while SIMnew simulates a weak carbon sink (−0.02 Pg C year−1) and SIMold simulates a stronger carbon sink (−0.13 Pg C year−1) (Fig. 3B). The bias in SIMnew is nearly within the sampling uncertainty (±118%) of the ML-based upscaling product, while the biases from both simulations fall within the large uncertainty (±1100%) in in situ average estimations (12). The differences between the two simulations could result from more source regions in Alaska and eastern Siberia in SIMnew than in SIMold (fig. S8, J and K) resulting from Yedoma carbon decomposition in SIMnew in these areas (Fig. 1D). Evaluation of simulated carbon fluxes across the ABZ and its two subregions against an ML-based upscaling product during the period of 2001 to 2014. (A to C) Multiyear mean GPP, Reco, and NEE from simulations using the old protocol (SIMold) and the new protocol (SIMnew), and the ML-based upscaling product, across the ABZ (A), Tundra ecosystem (B), and Boreal ecosystem (C). The positive NEE and negative NEE indicate net carbon source and net carbon sink, respectively. yr, year. (D to F) Trend in GPP, Reco, and NEE across the three regions. In each figure, both the annual fluxes (indicated as markers) and their two seasonal components including summer (light green bars) and nonsummer (light yellow bars) are shown. In (D) to (F), the bars with the outline and the markers filled indicate that the trends are statistically significant (P < 0.05). Given that the ABZ includes both tundra and boreal ecosystems, panels (B) and (C) are subsets of (A), and panels (E) and (F) are subsets of (D). The discrepancies in CO2 flux trends across the ABZ in SIMnew and SIMold during 2001 to 2014 against the ML-based product are more pronounced than the multiyear mean CO2 flux. The annual GPP and respiration trends are 1.4 and 2.4 times larger, respectively (Fig. 3, D to F, and fig. S9). This discrepancy may arise from the IPSL climate model’s warming rate during 2001 to 2014, which is nearly twice larger than the reanalysis dataset (ERA5) used for upscaling in the ML-based product (fig. S5, G, I, and K). When we normalized CO2 flux trends by the air temperature trend to assess the sensitivity of CO2 fluxes to air temperature, we found that the sensitivity of GPP and Reco to air temperature in the simulations closely matches that in the ML upscaling across all regions and seasons (table S2 and fig. S10), reinforcing our confidence in conducting future projections. For the NEE trend, the simulations are consistent with the ML upscaling during the nonsummer season, particularly in the tundra ecosystem. However, during summer and on an annual scale, our simulations show a smaller NEE sink trend or even different NEE trend directions compared to the ML upscaling. This discrepancy may result from the model’s lower sensitivity of GPP to air temperature (−60 to −30%) or its higher sensitivity of Reco to air temperature (+1 to +38%) relative to the ML upscaling (table S2). On the other hand, it is important to note that NEE trends in the ML upscaling product and our simulations are not statistically significant (P > 0.05) in most regions (fig. S10), which limits the extent to which historical trend comparisons can robustly constrain future projections. The ML-based product is also subject to uncertainties, including the limited number of measurement sites, especially those with long-term records, and the underestimation of strong CO2 fluxes resulting from the monthly aggregation of fluxes (12). Change in projected future TOC by 2100 Over the 21st century, the IPSL model predicts a continuous rise in air temperature across the northern region (>30°N), with regional warming from 4.0°C under the low-emission scenario (SSP1-2.6) to 11.0°C under the high-emission scenario (SSP5-8.5) relative to 1900, with even more pronounced warming at higher latitudes (fig. S4, A and B). This amplified warming relative to the historical period substantially heats the land surface and subsoil, resulting in a 45 to 84% [8.0 to 15.1 Mkm2 (million km2)] loss in permafrost extent in SIMnew and a 46 to 86% (8.2 to 15.4 Mkm2) loss in SIMold (fig. S11), which is consistent with previous studies (35). Under future warming, TOC in SIMold increases continuously by 87 Pg C until the end of the century under SSP1-2.6 (Fig. 2A). Even under the two high-emission scenarios (SSP3-7.0 and SSP5-8.5), TOC continues to rise until around 2065, after which it begins to decline, suggesting that the warming-induced increase in Reco surpasses the increase in GPP. However, because of the substantial TOC accumulation before 2065, SIMold still shows a net TOC gain of 42 to 65 Pg C relative to 1900 by the end of the century. SIMnew also projects a continuous increased TOC (41 Pg C) throughout the century under SSP1-2.6 and a reversal in TOC change under SSP3-7.0 and SSP5-8.5. However, the reversal occurs about 10 years earlier than SIMold by around 2055 (Fig. 2A). Given that the TOC accumulation rate before the reversal is lower in SIMnew than in SIMold and the TOC loss rate after the reversal is higher, by 2100, SIMnew shows a net TOC loss of 3 to 32 Pg C relative to 1900 under the two high-emission scenarios. This net TOC loss is entirely driven from Yedoma deposits, which experience a TOC loss of 30 to 32 Pg C (Fig. 2D). The temperature sensitivity and baseline turnover base rates of Yedoma soils after thaw have been validated against in situ CO2 flux measurements and laboratory incubation experiments (section S2). The conventional PFTs also experience a TOC reversal under both scenarios in both SIMnew and SIMold, but they maintain a net TOC gain or a weak net TOC loss (−6 to 46 Pg C) by 2100 (Fig. 2B). By contrast, the peat PFT shows net TOC gains in both SIMnew and SIMold (Fig. 2C). However, the peat PFT begins to lose carbon after 2060 in SIMnew under the two high-emission scenarios, whereas in SIMold, it continues accumulating carbon until the end of the century (Fig. 2C). This suggests the crucial role of the two deep carbon reservoirs in shaping the predicted response of northern soil carbon. Mechanisms for the higher TOC loss in SIMnew To explain the differences in predicted TOC changes between SIMnew and SIMold, we first examined the simulated carbon fluxes into (Carbonin) and out of (Carbonout) the soil, i.e., total litter carbon input and total soil respiration. From 1900 to 2100, both Carbonin and Carbonout are projected to increase substantially in both simulations because of enhanced net primary production under climate warming. However, Carbonout increases at a faster rate than Carbonin (figs. S12 and S13), leading to the TOC reversal shown in Fig. 2. When comparing SIMnew and SIMold, we found that the change in Carbonin relative to 1900 remains similar between the two simulations (within ±0.2 Pg C year−1) throughout the historical period and across all future scenarios, while the change in Carbonout diverges markedly, with a difference of 0.7 to 1 Pg C year−1 between SIMnew and SIMold under the two high-emission scenarios (Fig. 4A). The faster Carbonout increase in SIMnew reflects a larger decrease in turnover time of soil carbon pools (fig. S14), accelerating soil carbon losses. The large Carbonout difference is primarily driven by the Yedoma PFT (63 to 66%, 0.5 to 0.6 Pg C year−1) and is almost entirely distributed within the permafrost region (Fig. 4B). This suggests that enhanced soil carbon decomposition in deep carbon–rich areas results in considerably higher TOC loss in SIMnew. More deep carbon exposed by the thickening ALT contributes to higher carbon fluxes out of the soil in the simulation using the new protocol. (A and B) Difference in change of carbon fluxes into (Carbonin) and out of (Carbonout) the soil between the simulation using the new protocol (SIMnew) and that using the old protocol (SIMold) during the historical period (1900 to 2100) and across the three SSP scenarios by PFT (A) and by region (B). (C) Temporal variation in mean annual ALT from 1900 to 2100 from SIMold and SIMnew across the common simulated permafrost region in 1900. (D) Temporal variation in the difference in SOCALT between SIMnew and SIMold from 1900 to 2100 across the same permafrost region as in (C). In ORCHIDEE-MICT, the increase in soil carbon decomposition could be driven by an increased decomposition rate and/or changes in the soil carbon substrate, such as a greater quantity of available soil carbon or a shift toward more labile soil organic matter forms. As the most important environmental factor controlling the model’s soil carbon decomposition rate, the annual soil temperature shows negligible differences between the two simulations throughout the historical period and across the three future scenarios (fig. S11C). This indicates that temperature-driven differences in decomposition rates are unlikely to be the main driver of the higher soil carbon loss in SIMnew than in SIMold. To investigate the effect of differences in the quantity of available soil carbon, we examined the SOC within the active layer thickness (SOCALT) from the two simulations, given that model decomposition is inhibited when the soil is frozen (fig. S15). We found that the difference in SOCALT between SIMnew and SIMold remains positive, within 5 to 20 Pg C before 2060, but rises sharply afterward under the two high-emission scenarios (Fig. 4D). This rise could drive higher soil carbon decomposition and its acceleration after 2060 in SIMnew. On the basis of temporal variations of air temperature (fig. S4A) and active layer thickness (ALT) (Fig. 4C), we inferred that the sharp increase in SOCALT after 2060 in SIMnew is likely due to the accelerated thickening of the active layer (~0.2 m year−1), which increases the exposure of frozen deep carbon in SIMnew and amplifies the difference in SOCALT between the two simulations. The accelerated thickening of the active layer after 2060 at a rate of 0.2 to 0.4 m year−1 under SSP5-8.5 occurs across all PFT groups (fig. S16). As a result, under SSP5-8.5, the SOCALT for conventional PFTs and Yedoma PFT in SIMnew is 30 and 69 Pg C higher than in SIMold by 2100, respectively (fig. S17, A and C). For peat PFT, the difference in SOCALT between SIMnew and SIMold decreases by 24 Pg C despite the thickening of the active layer (fig. S17B), likely because the deep and passive peat carbon declines because of the shortening peat carbon accumulation in the new SOC accumulation protocol (Fig. 1 and figs. S2 and S3). Meanwhile, the difference in peat SOCALT for the active soil pool between SIMnew and SIMold rises sharply after 2060 (fig. S17D), potentially resulting in the higher TOC loss for peat PFT in SIMnew. To assess how the two protocols differ in terms of SOC quality (i.e., decomposability), we quantified the lability of soil organic matter as the fraction of total SOC allocated to active and slow pools (fa+s) in 1900, the starting year of the historical period in our model simulations. Across most northern regions, fa+s is 6% higher in SIMnew than in SIMold and increases with latitudes (Fig. 5, A and B). This may be due to the activation of the SOC insulation effect on soil thermal processes in the new protocol, which lowers soil temperature (section S1). The extra deep carbon is concentrated in high-latitude cold regions (Fig. 1), and the lower soil temperature there imposes stricter constraints on decomposition in the model than the low latitudes (fig. S15). Therefore, over thousands of years of SOC accumulation, the decomposition of active and slow SOC—driving the transition of carbon from active and slow pools to the passive pool—occurs more slowly in SIMnew, ultimately resulting in higher fa+s. The relationship between fa+s and soil temperature found here is consistent with previous studies and recent field measurements (fig. S18) (27, 36). Comparison in SOC quality between the simulations with the new and old protocols and its implication on the projection of future SOC response. (A) Spatial pattern of the difference in the fraction of active and slow SOC in 1900 between the simulation using the new protocol (SIMnew) and that using the old protocol (SIMold). (B) Difference in the fraction of active and slow SOC between SIMnew and SIMold across areas north of 30°N, 40°N, 50°N, 60°N, and 70°N. (C) Fraction of active and slow SOC for the four PFT groups from the two simulations. (D) Spatial pattern of the fraction of active and slow SOC in 1900 from SIMnew. (E) Latitudinal pattern of the fraction of active and slow SOC from SIMnew, as well as six CMIP6 models, which are categorized into two types according to their latitudinal patterns. The color of markers for SIMnew indicates the change in air temperature from 1900 to 2100 under SSP5-8.5. (F) Projected change in active, slow, and passive soil pools by SIMnew and two types of CMIP6 models. Across PFTs, we also found that the Yedoma PFT, which is located in colder regions, has a higher fa+s (Fig. 5C). While the peat PFT may not experience extremely low soil temperature, its waterlogged conditions lead to a very strong oxygen limitation on decomposition, even slowing the transition from the active pool to the slow pool (fig. S14). Because of the shortened peat SOC accumulation duration by 4800 years in SIMnew, the peat PFT shows a 5% higher active SOC fraction than in SIMold (Fig. 5C) despite their identical total SOC amount (Fig. 1C). Given the faster turnover rates of active and slow SOC pools (fig. S14), this higher fa+s in Yedoma and peatlands may also contribute to more rapid SOC decomposition under future warming in SIMnew. Implications on the projection of future SOC response To further investigate the role of SOC quality in determining future SOC responses, we examined the general spatial pattern of fa+s from SIMnew and found a consistent latitudinal trend: The higher the latitude, the larger the fa+s (Fig. 5, D and E). Given that future warming is larger at higher latitudes (Fig. 5E and fig. S4), these regions could experience a more rapid loss of soil carbon than lower latitudes. In addition, we analyzed six CMIP6 models that used the similar soil carbon module to ORCHIDEE-MICT and reported stocks of the three soil carbon pools (37). Unlike SIMnew, these six models exhibit two distinct trends: (i) type A, in which fa+s remains constant (51 ± 11%) with increasing latitude, and (ii) type B, in which fa+s decreases from 38 ± 6 to 8 ± 0.5% with increasing latitude (Fig. 5E and fig. S19A). Despite differences in initial soil carbon stocks (fig. S20) and predicted future warming (fig. S21) across models, both types predict a net SOC gain relative to 1900 (48 and 46 Pg C) by 2100 under SSP5-8.5, contrasting with the net SOC loss (−43 Pg C) in SIMnew (Fig. 5F and fig. S19B). Compared to SIMnew, all six CMIP6 models have a lower fa+s in higher latitudes, e.g., 50% in type A models and 10% in type B models, against 80% in SIMnew for areas north of 60°N. Given that passive carbon is highly resistant to decomposition, even under accelerated warming at higher latitudes, these models likely fail to capture the key processes driving permafrost carbon loss. To test this hypothesis, we conducted additional simulations for the period of 1900 to 2100, initiated using the same total soil carbon stock as SIMnew but distributed using gridded active, slow, and passive SOC fractions from one type A model, one type B model, and SIMold in 1900 (see Materials and Methods). Driven by the same climate forcing as SIMnew under SSP5-8.5, the run using SIMold’s SOC quality projects a similar net SOC loss to SIMnew (Fig. 6), suggesting that the SOC quality plays a minor role compared to soil quantity in driving higher SOC loss in SIMnew than in SIMold. By contrast, the runs using SOC quality from the CMIP6 models project a substantial net soil carbon gain, indicating that the inaccurate representation of SOC fractions in CMIP6 models may limit their ability to predict future SOC responses. Impact of soil quality on projected change in SOC during the period of 1900 to 2100 under SSP5-8.5. The four simulations used the same total amount of SOC from the simulation using the new protocol (SIMnew) but different SOC quality, i.e., the fraction of active, slow, and passive soil carbon distribution from original SIMnew, the simulation using the old protocol (SIMold), one type A CMIP6 model, and one type B CMIP6 model. DISCUSSION Representing deep carbon accumulation, including Yedoma deposition during the Pleistocene and peatland development during the Holocene, into the ORCHIDEE-MICT model increases simulated northern (>30°N) soil carbon stocks and shifts its soil carbon quality toward a more decomposable state. Compared to the original model version, the updated model shows better agreement with measurement-based estimates of soil carbon stocks [shown in the work of Xi et al. (31)], terrestrial ecosystem CO2 fluxes (Fig. 3), and soil carbon quality (fig. S18). Projections with this updated model indicate a substantial reduction in the previously simulated net sink of northern soil carbon between 1900 and 2100 by 47 to 74 Pg C across three SSP scenarios. Under the two high-emission scenarios, the incorporation of deep, old soil carbon deposits reverses the net cumulative carbon gain (43 to 65 Pg C) predicted by the original model to a net cumulative loss (−3 to −32 Pg C) and advances the transition of northern soil carbon balance from gain to loss predicted by the original model and other ESMs (18) to occur within this century. This higher soil carbon loss in the updated simulations is primarily driven by the gradual degradation of deep permafrost carbon, particularly from Yedoma deposits, which become increasingly exposed as the ALT deepens more rapidly after the mid-21st century. Inclusion of abrupt thaw processes, such as thermokarst lake formation, in the current model version would likely further exacerbate carbon release from deep permafrost. The greater proportion of easily decomposable permafrost carbon in the updated model also enhances soil decomposition. However, given that the shift in soil carbon quality is relatively small, its influence on future SOC dynamics is far less pronounced than the impact of increased soil carbon stocks. Nevertheless, when comparing CMIP6 models with both our simulations and in situ measurements, the substantial disagreements in the latitudinal distribution of soil carbon quality may help explain why current CMIP6 models predict a future increase in northern soil carbon stock. These findings are crucial for predicting future permafrost carbon-climate feedbacks, as they highlight the importance in including deep carbon dynamics when assessing the northern soil carbon balance. Our work also provides a valuable methodology for incorporating deep carbon processes into ESMs, which is still lacking in CMIP6 models. The methodology we implemented for deep carbon initialization in permafrost and cold regions in ORCHIDEE-MICT is applicable to other land surface models as well. Besides, because ORCHIDEE-MICT represents a single modeling framework, several uncertainties remain and require careful consideration. For example, it is crucial to further verify whether our model has accurately simulated the historical response of soil carbon to warming (as evaluated in Fig. 3) (10, 12, 38), whether the climate forcing used in our simulations realistically represents contemporary and future warming magnitudes (as evaluated in fig. S5) (18), and whether the absence of explicit microbial dynamics causes the model to overlook the microbial-mediated feedbacks on soil decomposition and it therefore may underestimate soil carbon loss (39, 40). Consequently, we encourage expanded observational coverage as well as similar model developments and test simulations by other modeling groups to comprehensively assess how deep carbon representation will ultimately influence the response of permafrost carbon decomposition to arctic warming in the future. As previously mentioned, beyond deep carbon, additional critical permafrost processes still need to be integrated into ESMs. One key issue is the representation of ground ice and the formation of thermokarst lakes resulting from ground-ice melting, which could further accelerate the decomposition of deep permafrost carbon (13, 17). If methane emissions from thermokarst lakes are also considered, northern regions could become an even larger source of greenhouse gas emissions than our simulations suggest (16, 41, 42). Another crucial issue is the interaction between wildfire and permafrost, especially given the increasing frequency and intensity of northern wildfire (12, 43). To what extent wildfire accelerates permafrost thaw and soil carbon decomposition—directly by heating or burning the soil and indirectly by altering surface properties such as albedo—remains a challenge for future model development (15). In addition, changes in northern vegetation dynamics with warming (44) and the release of reactive nitrogen as the permafrost thaw (45, 46) could also regulate soil carbon stocks. Improving the representation of high-latitude vegetation dynamics and incorporating key nutrient cycles into ESMs are also essential as we work to refine models’ ability to assess permafrost carbon-climate feedbacks. MATERIALS AND METHODS Soil carbon module description The soil carbon module of the original ORCHIDEE-MICT model (version 8.4.1) (29) is based on the CENTURY model (47), including two litter carbon pools (metabolic and structural) and three soil carbon pools (active, slow, and passive) defined by their turnover rates. The turnover times at 5°C without moisture limitation for the five carbon pools are 0.37, 1.4, 0.84, 31, and 1363 years, respectively. The variation in turnover time (or decomposition rate) with temperature is shown in fig. S15. During decomposition, litter carbon is partly released as CO2 and partly transferred into vertically resolved soil carbon pools; similarly, decomposed active SOC is partly respired as CO2 and partly transferred to slow and passive carbon pools, while decomposed slow SOC contributes to the passive carbon pool. The model does not include explicit microbial processes; thus, microbial-mediated feedbacks on decomposition are absent from our simulations. The vertical transfer and accumulation of soil carbon are driven by three key processes: root density–dependent litter carbon inputs, depth-dependent decomposition of soil carbon regulated by vertically layered soil temperature and soil moisture (including the pore ice effect), and soil carbon diffusion via bioturbation by animal and plant activity and via cryoturbation in permafrost soils. The model version used in this study assumes a total soil depth of 38 m (32 layers) for soil carbon processes. However, because of root depth limitations and ALT, the deepest simulated conventional soil carbon in the original ORCHIDEE-MICT model does not exceed 3 m (21). Later, Zhu et al. (48) introduced the sedimentation process into ORCHIDEE-MICT, enabling the model to simulate Yedoma deposits as deep as 20 m. Qiu et al. (49, 50) implemented an additional soil tile to host peatlands, allowing the model to simulate peat carbon accumulation up to 10 m. However, these advancements, developed in separate branches of ORCHIDEE-MICT, have yet to be integrated together for a realistic, process-based simulation of vertical and horizontal SOC distributions across the northern high latitudes. Implementation of old and new soil carbon accumulation protocols Given that the implementation of old (or “standard”) and new soil carbon accumulation protocols has been detailed in the work of Xi et al. (31), we provided only a brief overview in this study. The old protocol followed the classical spin-up simulation approach used in conventional land surface models to accumulate SOC stocks to match exactly the observations. It included three simulation stages: (i) 100 years with the full ORCHIDEE land surface model (FullO) run to initiate the energy cycle, hydrological processes, and vegetation growth in the model; (ii) 13,500 years with a stand-alone soil carbon module (SubC) run to accelerate carbon accumulation (using fixed monthly litter input, soil moisture, and soil thermal conditions from the preceding FullO simulation); and (iii) 100 years with the FullO run to bring energy, hydrology, and carbon back into equilibrium. The forcing datasets used in each stage are listed in table S1. The old protocol was conducted using ORCHIDEE-MICT version r8205, which included peat carbon accumulation but did not yet incorporate the sedimentation process forming Yedoma deposits (30). Besides, the old protocol applied a uniform carbon accumulation duration for all peatlands, therefore not accounting for the dynamic development of peatlands during the Holocene. In the new protocol, we first incorporated the sedimentation process to form Yedoma deposits into the reference ORCHIDEE-MICT version (r8205), following Zhu et al. (48). In reality, Yedoma deposits accumulated upward over time; in the model, this process was represented as an equivalent downward transport of SOC using a vertical advection equation. The downward, or sedimentation, rate of SOC was inferred from the height-age slope of Yedoma deposits obtained from in situ measurements (51). The Yedoma simulation was conducted under the Last Glacial Maximum climate conditions, with Last Glacial Maximum CO2 levels (185 parts per million) and the dynamically simulated vegetation cover (table S1). A map of present-day Yedoma from Strauss et al. (25), covering 0.48 Mkm2 of Yedoma deposits, was used to limit the model to simulate Yedoma deposits that have survived the deglaciation. The simulated mean deposit thickness (20 m) and TOC stock (141 Pg C) for present-day Yedoma agree well with the estimation based on site measurements (23, 52, 53). Radiocarbon dating of peat cores has revealed the heterogeneous inception times of global peatlands (54), indicating that prescribing a uniform age for all peatlands, as done in the old protocol, is unrealistic, given that peatland age can substantially affect peat carbon accumulation. To address this, we integrated radiocarbon dates from 2860 geolocated peat cores (55) and transformed them into gridded peatland cohort maps for each thousand-year interval since 16 ka [thousand years before present (1900)] using a gridded peatland fraction map (6, 31). Following the Yedoma carbon accumulation simulation, we conducted transient SOC accumulation simulations 1000 years at a time for peat PFT and 15 conventional PFTs (including one bare soil PFT, eight tree PFTs, four natural grass PFTs, and two crop PFTs) from 13.5 ka to the present (fig. S1 and table S1). As a result, compared to the old protocol, the new protocol simulates shallower peat carbon depths (by 1 to 5 m) and a reduction of 35 Pg C (43%) in passive soil carbon across northern peatlands resulting from the shorter accumulation duration of peat carbon. This reduction is particularly pronounced in younger peatlands [see the work of Xi et al. (31) for details]. Historical and future simulations following the old and new protocols On the basis of preindustrial soil carbon accumulated under the old and new protocols, we conducted historical (1900 to 2014) and future (2015 to 2100) simulations (fig. S1), including three SSP scenarios (SSP1-2.6, SSP3-7.0 and SSP5-8.5), using the bias-adjusted atmospheric climate data from the IPSL-CM6A-LR climate model within the Inter-Sectoral Impact Model Intercomparison Project 3b (table S1) (32). The historical climate was evaluated using the fifth generation ECMWF reanalysis dataset (ERA5) (32, 56) for their overlapping period (1979 to 2014) (fig. S5). The land cover maps was created by combining (i) 15 conventional PFTs from the TRENDY (trends in the land carbon cycle) project (57), consisting of one bare soil PFT, eight tree PFTs, four natural grass PFTs, and two crop PFTs; (ii) one Yedoma PFT based on the Yedoma map from Strauss et al. (25); and (iii) one peat PFT defined as peatlands that incepted before 0 ka (table S1) (6). The land cover maps were time-varying during the historical period but fixed for future projections using the 2015 land cover map. The fractions of Yedoma and peatland remained constant throughout both the historical period and three future scenarios. When creating the land cover maps, we prioritized Yedoma and peat fractions within each grid cell and proportionally adjusted the fractions of the 15 conventional PFTs from the TRENDY land cover map. The atmospheric CO2 concentration was time-varying throughout both the historical period and three future scenarios. From 2015 to 2100, the mean annual air temperature across the northern region (>30°N) is projected to increase by 1.5° to 8.4°C, annual precipitation by 28.4 to 94.3 mm year−1, and annual atmospheric CO2 concentration by 42.0 to 702.0 parts per million (fig. S4). All the simulations were run for the northern region (>30°N) at a 2°-by-2° spatial resolution. To evaluate the role of SOC quality in predicting future SOC responses, we conducted three additional simulations for the period of 1900 to 2100 under the SSP5-8.5 scenario (Fig. 6). These simulations followed the same setup as the historical and future simulations using the new protocol (SIMnew) but were initialized with the same total SOC stock as SIMnew but distributed using gridded active, slow, and passive SOC fractions from one type A model (CNRM-ESM2-1), one type B model (CESM2), and SIMold, respectively. Because of substantial disagreements between CMIP6 models and our simulations, the distribution scheme did not perform well for all grid cells. In cases where a specific SOC form was absent in the CMIP6 models but present in our simulation, we used the absolute value of that SOC form from SIMnew to initialize the SOC map. Measurements or measurement-based datasets To evaluate the simulated carbon cycle during the historical period, we used an ML-based upscaling product for terrestrial ecosystem CO2 fluxes including GPP, Reco, and NEE across the ABZ (12). To our knowledge, this ML-based product integrated the most extensive in situ CO2 flux measurements within the ABZ, covering 3968 to 4897 site-months, depending on the flux. Each site includes an average of 25 (ranging from 1 to 213) months of measurements, with long-term records exceeding 7 years available for seven boreal forest sites, one wetland site, and one tundra site. Given that only 14% of sites have been subject to natural or anthropogenic disturbances, the dataset primarily represents undisturbed ecosystem CO2 fluxes. Our simulations do not account for natural disturbances such as fire or anthropogenic activities, so the simulated CO2 fluxes can be directly evaluated using this dataset. The upscaling to the ABZ was conducted using random forest models for the period of 1990 to 2020 at a spatial resolution of 1 km by 1 km, but only data during 2001 to 2020 are available. While the upscaling results have been validated using various performance metrics against site-level measurements, they tend to underestimate strong sinks and sources because of the aggregation of fluxes at the monthly scale (12), which may partly explain the higher CO2 fluxes in our simulations compared to this dataset (Fig. 3). Moreover, for comparisons between simulations and the dataset across the ABZ’s tundra and boreal biomes, we used the same biome classification as used in the ML-based product from Dinerstein et al. (34). To evaluate latitudinal patterns in the fraction of active and slow SOC from our simulations and CMIP6 models, we used in situ SOC fraction measurements of particulate organic carbon (POC) and mineral-associated organic carbon (MAOC) concentrations from 133 sites across cold terrestrial ecosystems, including Arctic, subarctic, and alpine biomes (27). POC is free in the soil matrix, while MAOC is stabilized by binding to mineral surfaces. As a result, POC is more vulnerable to microbial decomposition under warming. Following previous studies, we compared the measured POC fractions with the simulated fractions of active and slow SOC (36). We acknowledge that this is a coarse approximation, as the three soil carbon pools in the CENTURY model do not fully capture the complexity of SOC dynamics in reality and cannot be directly equated to observed fractions. Nonetheless, this approach remains useful for assessing the dominance of active and slow SOC in cold regions (fig. S18), rather than making precise quantitative comparisons. CMIP6 model outputs The historical simulations and future projections under SSP5-8.5 of soil carbon (Figs. 5 and 6 and fig. S20) and air temperature (fig. S21) were obtained from the CMIP6 archive (37). We selected six CMIP6 models including ACCESS-ESM1-5, CESM2, CNRM-ESM2-1, IPSL-CM6A-LR, MIROC-ES2L, and NorESM2-LM, which used the same soil carbon module as ORCHIDEE-MICT and provided soil carbon stock data for all three SOC pools over the period of 1900 to 2100. For comparison with our simulation outputs, all model data were regridded to a 2°-by-2° resolution.