Climate forecasts project change not only in the mean of climate variables but also in their variance. If these dual changes interact, then future ecological dynamics will be difficult to predict using current experimental approaches, which typically change the mean or impose a single extreme event, such as drought. We designed a new field experiment to factorially reduce mean precipitation and increase its interannual variability. Across 4 years, drier, more variable precipitation additively reduced aboveground primary productivity by 48%-69% and interactively reduced the dominant plant species, but had no effect on the plant species predicted to dominate in the future, which could lead to state transition. Drier, more variable precipitation also interactively reduced biodiversity more than either climate factor alone, with 37%-42% fewer plant species than under ambient conditions, a pattern that matched declining richness during the past 20 years of ongoing climate change. Drier, more variable precipitation restructured the composition and spatiotemporal variation of the plant community. Altered precipitation mean or variance affected 14% of plant species, with eight species sensitive to the mean × variance interaction. Results suggest that future forecasts of plant community structure may be inadequate if they fail to incorporate climate mean × variance interactions.

Warming and elevated CO2 (eCO2) are two potentially opposing climate-carbon (C) feedback mechanisms that modulate the magnitude of the land C sink, with warming decreasing and eCO2 increasing C sequestration. However, their net effect on soil organic C (SOC)-the largest terrestrial C stock-remains uncertain. Here, we quantify how warming, eCO2, and their interactions influence SOC by using 5558 paired observations from 1392 global studies across ecosystem types. Our study shows that warming reduces SOC by 8.0%, primarily by suppressing aboveground C inputs (-1.0%) and decreasing microbial C use efficiency (-9.9%). Concurrent warming and eCO2 increase SOC by 7.6%-a synergistic effect larger than eCO2 alone (+4.8%), primarily contributed by croplands. This outcome may result from increases in plant C inputs and soil nitrogen availability under eCO2, which facilitate microbial C assimilation and necromass formation (+3.9%). These processes promote the accumulation of mineral-associated C (+9.9%) and offset the negative effects of warming. The combined effect of warming and eCO2 is projected to increase SOC by 27.4 Pg C by 2100. Our findings highlight that synergistic interaction between warming and eCO2 increases SOC sequestration and enhances SOC stability under future climate change.

The CO2-fertilisation effect (CFE) on vegetation productivity is the major driver of the enhanced land carbon sink in recent decades. CFE theoretically increases with elevation due to the higher sensitivity of carboxylation to an increase of CO2 under lower CO2 partial pressure, but the elevation-dependent CFE pattern has been largely overlooked. By conducting a 6-year CO2 enrichment experiment (+100 ppm) in an alpine grassland, we show that elevated CO2 increased gross primary production (GPP) by 25.5% ± 4.6%. Water availability and plant biomass allocation modulates CFE during different seasons. A global synthesis of 10 CO2 enrichment experiments reveals that CFE increased with elevation. The satellite-based EC-LUE model also demonstrates a positive global elevation-dependent CFE pattern, albeit substantially weaker than that from experimental observations. Current terrestrial biosphere models, however, could not represent the elevation-dependent pattern, highlighting the need to improve the representations of plants' elevational physiological adaptation to rising CO2 in models.

Plants often exhibit a midday depression in water use (i.e., transpiration), reflecting a constraint on their ability to sustain maximum water transport, which may occur at the cost of reduced photosynthesis. Eddy covariance observations and geostationary satellites cannot quantify this widespread phenomenon globally while resolving fine-scale spatial variability. Using evapotranspiration measurements from the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) and machine learning, we quantify the global distribution of midday depression in evapotranspiration. Midday depression primarily occurs during peak-growing seasons in temperate zones and dry periods in the tropics, with a morning shift of the peak evapotranspiration time by > 0.5 h. Droughts and heatwaves intensify midday depression, with peak evapotranspiration time shifting earlier, typically by 0.5 to 1.5 h and approaching 2 h during local extremes. By contrast, the coarse resolution upscaled flux products and hourly meteorological reanalysis datasets cannot capture this phenomenon. Midday depression of evapotranspiration detected using ECOSTRESS and machine learning is strongly associated with plant water stress, and helps us better understand how plants use water during drought and heatwaves.

ABSTRACT Soil microbial ecology is data‐rich yet struggles to predict ecosystem functions under ongoing environmental change, including climate warming, altered precipitation patterns, and land‐use intensification. The core challenge lies in linking micron‐scale microbial metabolism to field‐scale outcomes. We argue that progress requires: (i) shifting from taxonomic inventories to functional traits as predictors; (ii) integrating machine learning with ecological theory, rigorous uncertainty quantification, and cross‐validation; (iii) developing hybrid models that merge mechanistic understanding with data‐driven approaches; and (iv) building new theoretical and mathematical frameworks to bridge spatiotemporal scales. We distinguish soil function (ecosystem‐level outcomes) from microbial function (organismal or community‐level activities), as the relationship between them remains poorly understood. We focus on agricultural soils as a primary application domain, given their global significance for food security, climate mitigation, and the urgent need for management‐relevant predictions. Realising predictive microbiology demands standardised data, harmonised metadata, and transdisciplinary collaboration among soil scientists, ecologists, microbiologists, bioinformaticians, and data scientists. This commentary outlines a pathway from descriptive characterisation toward predictive, microbially informed management of multifunctional agricultural soils.

Global precipitation regimes have been shifted in recent decades, imposing significant consequences in water-limited grassland ecosystems. However, the effects of increased precipitation on the succession of soil microbial communities remain unclear, mainly due to the scarcity of long-term experiments with time-series data. Here, we examined temporal succession of grassland soil microbial communities in a long-term increased precipitation experiment. Both soil microbial taxonomic and functional structures were significantly altered by increased precipitation. Increased precipitation significantly decelerated the succession rates of soil microbial functional structure (i.e. time-decay relationships). Consistent with the increased microbial decomposition and heterotrophic respiration, the abundances of soil microbial carbon decomposition genes were markedly enhanced by increased precipitation. Furthermore, increased precipitation stimulated genes involved in nutrient cycling processes, potentially promoting plant growth. Collectively, the contributions of stochastic processes in shaping microbial communities were increased under increased precipitation, suggesting that microbial successional trajectories may shift toward multiple alternative states characterized by greater stochasticity under future altered precipitation regimes.

Plants employ three N acquisition strategies to support productivity: root uptake from soils, symbiotic fixation, and resorption from senescing leaves. However, their global variation and coordination remain poorly understood. Here, we compile 3,193 field observations and use machine learning to map the three strategies and quantify their relative contributions. We estimate that global nitrogen uptake, fixation, and resorption amount to 559.2 ± 56.5 (mean ± standard deviation), 125.5 ± 7.2, and 279.3 ± 5.0 Tg N yr-1, respectively, representing 58%, 13%, and 29% of total plant N acquisition. Their global variations are primarily determined by temperature and water conditions. Nitrogen uptake dominates globally, especially in warm, humid low-latitude regions, while resorption is more prevalent in arid and cold environments. In contrast, fixation contributes relatively little, with some hotspots in tropical regions and northern high latitudes. Notably, we find significant shifts in N acquisition strategies-from uptake to fixation and resorption-along gradients of decreasing nitrogen availability. These results suggest a spatially coordinated adjustment of plant nitrogen acquisition strategies in response to changing N conditions and offer insights into how coordinated nitrogen strategies shape terrestrial nitrogen cycling and its coupling with the carbon sink.

Soils have potential to mitigate climate change by sequestering carbon from the atmosphere, but the soil carbon cycle remains poorly understood. Scientists have developed process-based models of the soil carbon cycle based on existing knowledge, but they contain numerous unknown parameters and often fit observations poorly. On the other hand, neural networks can learn patterns from data, but do not respect known scientific laws, and are too opaque to reveal novel scientific relationships. We thus propose Scientifically-Interpretable Reasoning Network (ScIReN), a fully-transparent framework that combines interpretable neural and process-based reasoning. An interpretable encoder predicts scientifically-meaningful latent parameters, which are then passed through a differentiable process-based decoder to predict labeled output variables. While the process-based decoder enforces existing scientific knowledge, the encoder leverages Kolmogorov-Arnold networks (KANs) to reveal interpretable relationships between input features and latent parameters, using novel smoothness penalties to balance expressivity and simplicity. ScIReN also introduces a novel hard-sigmoid constraint layer to restrict latent parameters to prior ranges while maintaining interpretability. We apply ScIReN on two tasks: simulating the flow of organic carbon through soils, and modeling ecosystem respiration from plants. In both tasks, ScIReN outperforms or matches black-box models in predictive accuracy while greatly improving scientific interpretability -- it can infer latent scientific mechanisms and their relationships with input features.

The terrestrial carbon sink has strengthened in recent decades through enhanced maximum CO2 uptake (MCU) and a longer carbon uptake period (CUP), whereas maximum CO2 release (MCR) remains less quantified. Using atmospheric inversion estimates and 16 terrestrial biosphere models, we show that MCU and MCR both increased significantly from 1993 to 2022, with MCU rising faster globally (0.18 ± 0.04 versus 0.07 ± 0.01 g C m−2 yr−1). MCU increases were widespread, whereas MCR increases were strongest at mid- to high latitudes. Enhanced MCU and CUP increased annual net ecosystem productivity by 0.31 ± 0.09 g C m−2 yr−1, but rising MCR offset 44 ± 20% of this gain, leaving a net increase of 0.18 ± 0.06 g C m−2 yr−1. Elevated CO2 dominates long-term increases, while climate variability contributes to spatial heterogeneity. These findings highlight the importance of accounting for carbon release in projections of future terrestrial sink strength. This study shows that stronger maximum CO2 uptake and a longer carbon uptake period strengthen the terrestrial carbon sink but increasing maximum CO2 release offsets 44% of this gain, highlighting an important constraint on its future strength.

Organo–mineral associations are central to the persistence of soil organic carbon (SOC) in forest mineral soils, yet how nitrogen (N) enrichment affects mineral-associated stabilization across contrasting soil chemistries remains unclear. Using 1,274 paired observations from field N addition experiments spanning a broad soil pH gradient, we found that forest SOC responses to N addition were strongly pH dependent and followed a distinct non-linear (U-shaped) pattern. Nitrogen addition increased SOC in strongly acidic (pH < 4.5) and alkaline (pH > 7.5) soils, but had little effect at intermediate pH (4.5–7.5). Where SOC fractions were available, these contrasting responses were driven primarily by changes in mineral-associated organic carbon (MAOC), whereas particulate organic carbon showed a weaker contribution. Patterns of exchangeable cations supported mineral-mediated stabilization mechanisms, with greater MAOC accumulation associated with increased aluminium availability in acidic soils and increased base cation availability in alkaline soils. Model-based upscaling further suggested that spatial variation in soil pH may explain regional heterogeneity in SOC responses to contemporary N deposition, with stronger positive responses predicted for many tropical forests than for temperate and boreal forests. Together, these results identify soil pH as a key control over forest SOC responses to N enrichment through mineral-associated stabilization, and provide a process-based framework for predicting how anthropogenic N inputs may alter soil C persistence.

Plants assimilate carbon through photosynthesis (gross primary productivity, GPP) while losing water via transpiration (Trans), with both processes responding nonlinearly to temperature. Although the air temperature optimum of GPP ( T opt GPP ) is well studied, the thermal response of Trans ( T opt Trans ) remains unknown. Here, using global eddy covariance observations and sap flow measurements along with simulations from an Earth system model, we find that T opt Trans is consistently higher than T opt GPP across biomes and climates, indicating greater heat tolerance in Trans. Despite a strong correlation, their divergence suggests carbon uptake is more vulnerable to warming than water loss. Machine learning identifies maximum air temperature as the key driver of both optima, while their difference ( Δ T opt ) is associated with vegetation water content. The Earth system model predicts spatial patterns of T opt Trans and T opt GPP that align with observations, but the model significantly underestimates the magnitudes of T opt Trans , T opt GPP and Δ T opt . These results reveal a critical decoupling of carbon-water coordination under heat stress, with ecosystems sustaining Trans beyond T opt GPP to cool leaves, but ultimately reducing Trans to conserve water.

Ecology is entering a data-rich era driven by rapid advances in sensor technologies and the expansion of environmental observations across spatial and temporal scales. In parallel, machine learning (ML) has been increasingly adopted as a powerful tool for analysing complex ecological systems, where interactions among organisms and their environments are often nonlinear, heterogeneous and scale-dependent. Over the past three decades, ML applications in ecology and global biology have expanded rapidly, supporting advances in pattern detection, prediction, monitoring and process understanding, while still facing major challenges in data quality, sampling bias, interpretability, uncertainty quantification and causal inference. Emerging developments in embodied Artificial Intelligence (AI), foundation models, large language models and intelligent agent systems now create new opportunities for ecological research by enabling adaptive observation, improved integration of heterogeneous data and more efficient analytical workflows. These advances are beginning to steer the field towards an AI-assisted research paradigm, in which ecologists supervise coordinated AI systems for data collection, integration, analysis and modelling, while providing interpretation, validation and iterative feedback. Realizing the full potential of such human-AI synergies in ecological studies will require closer integration between ecological knowledge and advanced AI methods, as well as sustained collaboration among ecologists, data scientists and technology developers.

Anthropogenic nitrogen (N) deposition is a major nutrient input to forests, yet most N addition experiments fall short of mirroring the chronic canopy‐level deposition observed in nature. As a result, we still lack an understanding of how atmospheric N deposition affects overall forest multifunctionality, including trade-offs and synergies among different ecosystem functions. Failure to account for these interactions risks that nutrient-related management strategies unintentionally undermine important functions while promoting others. Here, we integrate national forest inventory data with forest functioning models in a multivariate Bayesian framework to quantify the effect of N deposition on forest multifunctionality derived from 13 individual functions, explicitly accounting for their trade-offs. We identify a consistent negative effect of N deposition on biodiversity conservation-oriented multifunctionality (~95% probability) in both broadleaf and needleleaf forests. Biomass production-oriented multifunctionality shows a positive association with N deposition in broadleaf forests (73% probability), but neutral-to-negative responses in needleleaf forests. Overall and climate regulation-oriented forest multifunctionalities are both likely to decline with increasing N deposition in both forest types (67% – 81% probabilities). Such relationships are largely stable across temperature, precipitation, soil pH, and altitudinal gradients, suggesting broad applicability across climatically analogous temperate (ca. 51.6% of global area) and boreal (ca. 24.4% of global area) forests worldwide. Pervasive trade-offs among individual functions underscore the risk of ignoring inter-dependence that substantially biases estimates of N deposition effects on forest functioning. Our findings highlight the importance of reducing nitrogen emissions, particularly for mitigating biodiversity loss for multifunctional forests.

Forest-based carbon dioxide removal (CDR) is widely promoted as a high-potential, nature-based climate solution, and its efficacy is time dependent because forest CDR follows the growth trajectory of aboveground biomass (AGB). However, this timing is rarely accounted for in CDR assessments, obscuring whether forest restoration can deliver meaningful CDR within 25–55-year climate horizons. Here we provide 1-km resolution, time-dependent estimates of forest CDR across the conterminous US, derived from 113,806 field-based AGB measurements from 74,048 plots. We show that the window for effective forest-based CDR is limited by age-related declines in tree growth. Forest CDR via AGB growth peaks within 9.4–31.2 years, then weakens as forests mature, falling to 1 MgCO2 ha⁻¹ yr⁻¹ within 51.6–108.5 years. Rather than emphasizing full CDR potential, we quantify a more attainable target, achieving 50% CDR potential. This target is reached fastest in the East South Central region within 18.9 ± 5.6 years and slowest in the Mountain region within 75.7 ± 23.9 years. Ignoring time dependence causes IPCC to underestimate CDR in young forests and overestimate it in older ones. Because forest carbon sinks take decades to develop, early action is crucial for restoration to deliver meaningful removals on near-term climate timescales.

Terrestrial ecosystems have cumulatively sequestered 24% of anthropogenic carbon dioxide (CO2) emissions since 1850 and are critical for mitigating future climate change. However, current Earth System Models (ESMs) remain highly uncertain in projecting future trajectories of this carbon sink capacity, hampering our predictive understanding of climate mitigation potential and impeding effective climate and carbon management policies. This study develops a novel framework that harnesses deep-learning (DL) to constrain uncertainties of ESM-projected Gross Primary Production (GPP) and Net Ecosystem Production (NEP) through 2100. Specifically, we apply DL to characterize the “offset” between ESM-simulated output (using CMIP6 models) and best-available observational products (top-down, bottom-up). This offset is treated as unresolved processes by current ESMs that could be effectively resolved by DL, which, once trained during historical periods, can be applied to adjust CMIP6 projections of the future. We find that DL significantly reduces the inter-model spread of GPP by ~56% and NEP by ~66% across the CMIP6 ESM ensemble . Under the medium emission scenario (SSP 245), the ensemble mean for NEP in 2100 is much weaker, 2.42 ± 1.16 PgC yr⁻¹ compared to 5.52 ± 3.45 PgC yr⁻¹ in the raw CMIP6 projections, suggesting a current overestimation of future carbon sequestration capability. Interestingly, DL revealed a slower trajectory of NEP growth compared to the raw CMIP6 projection. Beyond curbing the uncertainties of CMIP6 projections, DL also captures key environmental sensitivities of carbon cycle processes such as CO2 fertilization and sensitivity to warming. These findings demonstrate the power of DL in effectively curbing ESMs projection uncertainties and suggest that relying solely on natural terrestrial carbon sinks for climate mitigation is unlikely to slow down climate warming.

Non-structural carbohydrates (NSCs) serve as key metabolic substrates and carbon reserves in plants, playing a central role in regulating vegetation carbon dynamics. Despite their importance, the global distribution and underlying drivers of NSC variations remain uncertain. We compiled a database consisting of 29,386 field measurements on NSC concentrations during the growing season, covering 1,016 sites and 2,041 species, to examine organ-specific NSC patterns and disentangle the relative contributions of environmental conditions and evolutionary history. Leaf NSC concentrations increase with latitude, whereas stem and root NSC concentrations decline, linked primarily to water availability, temperature and solar radiation. The spatial contrast reflects differentiated source-sink controls on NSC storage along environmental gradients. In contrast, NSC concentrations increase with evolutionary divergence time in leaves, but decline in stems and roots. Crucially, evolutionary history explains a larger portion (55.9-77.1%) of global NSC variations across organs, surpassing contemporary environmental controls. These findings reveal the organ-specific disparity in NSC storage and underscore the central role of evolutionary history in shaping global NSC variability, with important implications for plant carbon balance and vegetation carbon modelling.

Human-driven increases in atmospheric CO2 (eCO2) are stimulating plant growth, thereby increasing the input of plant-derived carbon into soils. The fate of this additional carbon depends on the capacity of soil microbiomes to decompose and transform organic matter, a central process in regulating soil organic carbon (SOC) dynamics. However, how eCO2 affects this microbial capacity remains poorly understood. Because soil extracellular enzymes catalyse the degradation of various SOC pools, their activities (extracellular enzyme activities, EEAs) could offer mechanistic insights into microbially mediated SOC dynamics. We synthesized 272 observations on SOC and EEAs from eCO2 experiments across farmland, forest, grassland and shrubland, combining classical meta-analysis with random forest modelling. Our results showed that eCO2 significantly increased SOC by 4.2%. Among all variables tested, increased cellulase activity, which targets the breakdown of labile carbon sources, emerged as the strongest predictor of SOC accumulation. Specifically, eCO2 stimulated cellulase activity by 12.2% but had no effect on ligninase activity, which decomposes recalcitrant carbon. This enzymatic shift was likely driven by increased plant-derived labile carbon inputs under eCO2 and was associated with changes in the soil microbiome, including a higher fungi-to-bacteria ratio. These results underscore the potential of EEA as a predictive indicator of SOC accumulation under eCO2 and the importance of representing enzymatic processes in Earth system models.Read the free for this article on the Journal blog.

Natural disturbances are major drivers of large-scale forest dynamics. However, the representation of forest disturbances remains oversimplified and poorly constrained in terrestrial biosphere models, compromising predictive performance under rapidly changing disturbance regimes. In this review, we first summarize the general mechanisms underlying vegetation responses to major disturbance agents across spatio-temporal scales, including occurrence regimes, immediate structural impacts, post-disturbance community reorganization, and long-term ecosystem feedbacks. Predictable patterns of the processes provide the ecological foundation of disturbance modeling. Subsequently, we synthesize progress and challenges toward more mechanistic disturbance modeling, organized by three generic modules: occurrence, impact, and post-disturbance dynamics. Future model developments should implement dynamic disturbance regimes, model lethal and nonlethal structural damage with trait-based hazard functions, and represent key ecological processes determining post-disturbance ecosystem dynamics - including different regeneration strategies, trait plasticity in response to rapid microenvironmental changes, and disturbance legacy effects. Finally, we discuss observational constraints on mechanistic disturbance models. Model parameterization and development can benefit from advancing disturbance detection and attribution via multi-modal and multi-platform remote sensing, ground measurements, and disturbance manipulation experiments. Altogether, mechanistic, scalable, and observation-constrained simulation of disturbance is essential for reducing uncertainties in forecasting global change impacts on ecosystems and carbon cycle feedbacks.