Winter warming is altering plant exposure to cold events, making it increasingly important to understand how cold hardiness dynamics respond across the dormant season. Using the Spruce and Peatland Responses Under Changing Environments (SPRUCE) experiment in northern Minnesota, we measured bud cold hardiness across four dormant seasons (2021-2025). There, a boreal peatland community was exposed to continuous active whole ecosystem warming across five levels (+0.00°C to +9.00°C) combined with atmospheric and elevated CO2 (+500 ppm) in open-top chambers. Cold hardiness of two overstory (Larix laricina and Picea mariana) and two understory (Chamaedaphne calyculata and Rhododendron groenlandicum) species was evaluated semi-regularly throughout seasons. Both understory and overstory species gained cold hardiness (acclimated) at warmer temperatures in the fall than those causing deacclimation in the spring, demonstrating different temperature responsiveness phases across the dormant season. Based on cold hardiness sensitivity, warming did not alter maximum midwinter cold hardiness (sensitivity~0°C change in cold hardiness/°C of warming) but delayed fall acclimation and, most importantly, accelerated spring deacclimation (sensitivity > 0°C). These changes resulted in increased cold damage risk in late winter and spring with warming (sensitivity > 1°C/°C) for L. laricina, C. calyculata and R. groenlandicum. Among overstory species, the deciduous L. laricina was more sensitive to warming than the evergreen P. mariana, indicating warming may disadvantage acquisitive species. For understory shrubs, reduced snow cover in the warmest enclosures that removed the insulating buffer and exposed buds to temperatures colder than in control plots associated with the high sensitivity of these species resulted in observed midwinter bud mortality with warming. Rather than uniformly changing cold damage risk, warming decreases risk of cold damage in fall and early winter for all species, but increases risk in late winter and spring for most species, suggesting species-specific cold hardiness dynamics and snow-driven microclimate may result in boreal forest composition changes.

Boreal peatlands store 13 %-32 % of the global soil carbon (C) stock, a service dependent on plant-mycorrhizal fungi associations. In these nutrient poor systems, ectomycorrhizal and ericoid mycorrhizal fungi supply up to >80 % of the nutrient requirements of their plant hosts, partly with mined nitrogen (N) and phosphorus (P) from soil organic matter that are otherwise inaccessible to plants. Despite the ecological significance, mycorrhizal associations are only represented in a few land surface or ecosystem models. We modify the peatland branch of version 2 of the Energy Exascale Earth System Land Model (ELMv2-SPRUCE) to replace the default photosynthesis-driven inorganic N and P (NP) uptake process with a more realistic representation of the process via three pathways: (1) direct inorganic NP uptake by uncolonized fine roots, (2) indirect inorganic NP acquisition and (3) indirect NP acquisition from organic sources by mycorrhizal roots. We systematically evaluated the performance of the default and modified models with field observations from a whole ecosystem warming and carbon dioxide fertilization experimental site: Spruce and Peatland Responses Under Changing Environment (SPRUCE), in northern Minnesota, USA. The modified model reduces the underestimation of the growth response of shrubs in the default model to warming from 40 %-80 % to 17 %-35 % and reduces the overall relative absolute error on C fluxes from 1.61 to 1.54 in calibration. Improvements on modeled shrub growths and shrub-moss community net ecosystem exchanges are also seen in validation. The improved growth response of shrubs to warming is accompanied by several-fold increase in direct inorganic NP uptake and decrease in fungal colonization rate. The modified model simulates a smaller magnitude of transition of the ecosystem from C sink to C source under warming due to alleviation of plant nutrient limitation. Equifinality analysis shows the newly added parameters in the modified model can be constrained by the observed C fluxes. Sensitivity analysis shows the newly added parameters have stronger statistical interactions than the preexisting parameters in the default model. Overall, the modified model is an improvement over the default ELMv2-SPRUCE and will be a useful tool for understanding boreal peatland change.

Abstract The viability of peatlands as terrestrial carbon sinks is unknown as peatland carbon losses could be offset, or exacerbated, by increased plant growth. The response of fine‐roots to changing conditions in these ecosystems will be critical to understanding plant access to water, acquisition of nutrients, and ecosystem structure and function. We examined the plant functional type‐specific responses of fine‐root production to warming and elevated CO 2 treatments, and to warming treatment‐induced deepening of water‐tables in a forested boreal bog. We hypothesized that shrubs would show the most dynamic responses of fine‐root production. We used minirhizotron cameras to estimate fine‐root production, depth and peak annual standing crop from images of individually tracked roots across 7 years (2015–2021) in the SPRUCE (Spruce and Peatland Responses Under Changing Environments) experiment. Fine‐root production increased with warming for shrubs 10.3, trees 2.2 and herbs 4.6%°C −1 . When water‐tables dropped to their lows of 78 cm deep, shrubs deepened their fine‐root production from 7 to 29 cm deep, and trees from 9 to 17 cm, while herbs produced fine‐roots around 31 cm deep independent of shifts in water‐tables. Fine‐root standing crop increased with warming most strongly for shrubs, even more so under elevated CO 2 . Synthesis. We found that the fine‐roots of shrubs responded the most strongly to warming, elevated CO 2 and lowered water‐tables. The greater responsiveness of shrub fine‐root production may partially explain increasing ‘shrubification’ at this site and in other peatlands with long‐term drainage.

Abstract. Field-warming experiments offer insight into the response of ecosystems to rising temperatures, but cross-site comparison is needed to determine both the general tendencies of warming responses and the context dependencies of deviations from those norms. These responses are not limited to the direct effects of temperature but also their indirect effects on soil moisture, a critical factor controlling ecosystem productivity and carbon fluxes. Here we introduce SWEDDIE: the first database to characterize the whole soil profile warming response across 26 distinct soil warming experiments, encompassing forest, grassland, cropland, tundra, and wetland ecosystems. SWEDDIE is needed because prior databases and syntheses of warming effects on ecosystems were dominated by aboveground warming studies, many of which warmed soil modestly or negligibly during much of the growing season and reported only growing season averages. We demonstrate the potential of the SWEDDIE database by quantifying soil temperature and moisture changes for each experiment as a function of depth, warming methodology, ambient climate conditions, and ecosystem, as well as the relationship between soil moisture and imposed warming. Warming attenuated with depth at sites with aboveground warming only but increased with depth at sites with belowground warming only, as hypothesized. Warming led to soil drying at most sites, and drying was positively correlated with the magnitude of warming. However, the relationship between soil warming and soil drying varied by ecosystem: forest soils dried the most, while tundra soils became wetter with warming. Ambient climatic conditions also significantly influenced the relationship between experimental warming and drying, with more drying per degree of warming observed in soils with higher ambient moisture. The inconsistency of soil moisture changes with warming across ecosystems and warming methodologies demonstrates the importance of quantifying shifts in temperature and moisture in both space and time in order to overcome site-specific bias in ecosystem warming responses. The high temporal resolution and depth-resolved observations of the fundamental ecosystem properties of soil temperature and moisture in SWEDDIE v1.0.0 serve as a foundation for future experimental soil warming synthesis efforts and demonstrate the power of this actively growing community resource.

Leaf phenology may influence the development of wood structure and hydraulic function across growing seasons, yet the roles of green-up, green-down, and growing season length in regulating xylem anatomy remain unclear. We quantified annual wood anatomy, leaf phenology, and growth in a whole-ecosystem experiment with 5 warming levels (up to +9 °C) and 2 CO2 levels (ambient and +500 ppm) in Picea mariana (conservative spruce) and Larix laricina (acquisitive larch). We identified a phenology-tracheid-growth spectrum, reflecting a trade-off between hydraulic safety (thicker walls, higher tracheid density, and later green-up) and fast growth (wider tracheids, delayed green-down, and longer growing seasons). In spruce, earlier green-up, longer growing seasons, and later green-down increased the latewood hydraulic diameter more than the earlywood. In larch, earlier green-up increased earlywood hydraulic diameter, while later green-down increased latewood mechanical safety via thicker walls. Larch exhibited greater phenological sensitivity to elevated CO2 in regulating wood anatomy than spruce. Warming indirectly increased spruce growth by extending the growing season, which increased the latewood hydraulic diameter and subsequently enhanced overall growth. Warming directly increased larch growth but not through enhanced earlywood hydraulic conductivity. These findings demonstrate the role of divergent phenological adjustments in hydraulic function, with implications for boreal carbon and water fluxes.

Abstract. Peatlands cover only ~3 % of Earth’s land surface yet store ~30 % of global soil carbon (C), making them critical components of the terrestrial C cycle and influential regulators of C-climate feedbacks. However, their responses to climate warming and elevated CO₂ remain highly uncertain, in part because Earth system models represent peatland processes with varying levels of complexity and realism, and because peat C pools turn over on centennial to millennial timescales that challenge model evaluation. Here, we present results from the SPRUCE (Spruce and Peatland Responses Under Changing Environments) Model Intercomparison Project (SPRUCE-MIP), which evaluates 15 terrestrial ecosystem models against observations from a long-term whole-ecosystem warming and CO₂-enrichment experiment at an ombrotrophic bog in northern Minnesota, USA. Models were driven by observed meteorology from in situ warming treatments spanning +0 to +9 °C, at either ambient or elevated CO₂ (+500 ppm). Simulated net ecosystem exchange (NEE), net primary productivity (NPP), heterotrophic respiration (HR), and methane (CH₄) fluxes were benchmarked against multi-year observations. Model predictions exhibit a large spread in both baseline C balance, ranging from strong sinks to strong sources under +0 °C warming, and temperature sensitivity, indicating substantial uncertainty in predicting peatland C responses to environmental forcing. Across the ensemble, models fall into four distinct functional response types: (i) models that simulate peatlands as net C sources across all warming and CO₂ conditions; (ii) models that transition from C sinks to sources under warming irrespective of CO₂ level; (iii) models that transition only under ambient CO₂ but remain C sinks under elevated CO₂, reflecting strong CO₂ fertilization effects; and (iv) models that maintain persistent sinks or near-neutral C balance even under extreme warming. While most models predict enhanced NPP under elevated CO₂ across all warming treatments, SPRUCE observations show little or no NPP enhancement under elevated CO₂ at the +0 and +2.25 °C warming levels, with a positive CO₂ fertilization effect emerging only under stronger warming. This discrepancy highlights persistent model biases in representing interactions between warming and CO₂ responses. Process-based analysis further indicates that divergence in modeled responses is associated with differences in vegetation structure (light competition and moss representation), nutrient cycling (nitrogen and phosphorus), and dynamic peat representation. These structural differences systematically influence ecosystem productivity, heterotrophic respiration, and net C balance across the model ensemble. An illustrative parameter sensitivity analysis using the SPRUCE-specific Energy Exascale Earth System Model (E3SM) Land Model (ELM-SPRUCE) further shows that parameter choices can also modulate the magnitude of simulated responses. Together, these results demonstrate that peatland responses to warming and elevated CO2 are highly sensitive to model structure, parameterization, and process coupling, highlighting the need for improved representation of key peatland processes to reduce uncertainty in Earth system projections. The SPRUCE experimental framework provides a unique benchmark for evaluation and improving process representation and constraining near-term peatland response to environmental change, thereby strengthening confidence in longer-term C-climate projections.

Boreal bogs sequester large stores of terrestrial carbon in waterlogged peat occupied by extensive networks of vascular plant roots. Historically, cold temperatures and shallow water tables in bogs have favored the growth of Sphagnum over vascular plants. Seasonally variable water tables in bogs constrain woody shrub and tree root production to shallow, oxic acrotelm peat horizons while herbs with aerenchymous root tissue can grow below the water table. Warming and elevated CO2 are altering ecosystem functioning in nutrient-limited, rain-fed (ombrotrophic) bogs, via direct impacts on physiology and indirectly via altering water table levels. The environmental changes are shifting interactions between vascular plant fine-roots and the surrounding peat horizons and increasingly favoring vascular plant growth relative to Sphagnum sp. Altered functioning of fine-roots may dramatically affect the role of plants in these ecosystems because of the importance of these organs for plant resource acquisition. We examined how fine-roots across vascular PFTs differentially respond to warming & elevated CO2 manipulations and the consequent water table depression in a forested, boreal bog. We used minirhizotrons (cameras inserted belowground) to measure fine-root production, depth distribution, and standing crop from 2015-2021 for each PFT. We found that daily rates of fine-root production accelerated more for shrubs than for trees or herbs. Shrubs and trees grew their roots more deeply with depressed water table levels and shrub fine-roots became narrower. On an annual basis, fine-root production increased with warming, though these rates varied among PFTs and exhibited interactions with elevated CO2. Standing crop of fine-roots increased with warming temperatures, most strongly for shrubs under elevated CO2, because of these productivity responses. From these results, we expect boreal bogs will become increasingly dominated by shrubs under future warmer temperatures, higher atmospheric CO2 concentrations, and lower water table levels. ### Competing Interest Statement The authors have declared no competing interest. United States Department of Energy, https://ror.org/01bj3aw27, Terrestrial Ecosystem Science Scientific Focus Area

Global warming and increasing air temperatures also result in rising soil temperatures. Although acceleration of soil organic carbon cycling can be expected, the order of magnitude and speed of adaptation of carbon cycling to warming still remains largely unknown. This is especially crucial in boreal peatlands, where large reserves of terrestrial carbon are stored and these systems are known for their vulnerability to environmental changes. We investigated the organic matter composition in the SPRUCE (Spruce and Peatland Responses Under Changing Environments) experiment, where a boreal peatland was exposed to temperatures of up to +9°C and increased CO2 concentration compared to control conditions in open top chambers. A broad set of molecular markers (e.g., free extractable and bound lipids, lignin, benzene polycarboxylic acids) was used to trace incorporation and cycling of organic matter in the peat profile down to three meters depth four years after the start of the experiment. A strong response to increasing temperature was observed in the plant, microbial and peat chemical composition, the latter mainly in the acrotelm (0-30 cm) and partially also in the mesotelm (30-70cm). The response of the plant chemical composition was species-specific with the exception of nitrogen concentrations that increased for all plants. This is related to the stronger degradation of peat organic matter and thus increasing availability of nitrogen with rising temperature. All investigated molecular markers indicated a very fast response of carbon cycling in the whole acrotelm of the peat profile. This resulted from a dropping water table and thus more oxic conditions in the peat, which further enabled increasing shrub and tree root growth and increasing microbial abundance and activity. As a consequence of the more aerobic conditions, not only the comparatively easily degradable free extractable lipids, but also slow cycling polymeric substances such as suberin/cutin, lignin, and benzene polycarboxylic acids rapidly degraded and reflect an unexpectedly fast cycling of organic matter in the boreal peatland with increasing temperature. The acceleration of carbon cycling within the peatland with rising temperature is also reflected by the partial uptake of respired CO2 by the plants as indicated by the bulk and compound-specific d13C composition of the plants. Overall, our results illustrate the fast alteration of organic matter cycling in a boreal peatland when exposed to increasing temperature.

The Spruce and Peatland Responses Under Changing Environments (SPRUCE) experiment has operated five whole-ecosystem warming manipulations (+0, +2.25, +4.5, +6.75, and +9 degrees C) with paired ambient and elevated CO2 atmospheres (eCO2, +500 ppm) for 8 full calendar years (since August 2015). We tracked shrub-layer vegetation responses to the treatments using annual destructive plot sampling. Tree (Picea and Larix) responses were assessed annually using nondestructive dimensional analyses and allometric conversions. Shrub community changes were assessed for key ericaceous shrubs (Rhododendron, Chamaedaphne, and Kalmia), two Vaccinium species (V. angustifolium, V. oxycoccos), graminoid species (mostly Eriophorum), and one common forb (Maianthemum trifolium), plus minor understory species. We tracked annual aboveground net primary production (ANPP) for vascular plant species in gC m-2 y-1 and overall stand contribution in dry mass. We observed a linear increase in shrub-layer aboveground biomass accumulation with warming over time due primarily to an increase in ericaceous shrub abundance. Cumulative biomass increases across the shrub community showed overall positive responses to eCO2 after 8 years. Community composition also changed with warming, with increases in woody shrub density, and the reduction or loss of forbs. The tree community showed minimal initial responses to warming early in the treatments, but since 2020, has shown significant increases in ANPP and individual tree growth with warming. The main driver of change in the vascular plant community was temperature, with less pronounced effects of eCO2 evident. These results indicate an overall increase in ANPP with warming from both the tree and shrub layers of peatland vegetation.

Extreme drought events are predicted to increase with climate change, yet their impacts on ecosystem carbon dynamics under warming and elevated carbon dioxide (eCO 2 ) remain unclear. In a peatland experiment with five warming treatments each under ambient carbon dioxide (aCO 2 ) and eCO 2 (+500 parts per million), a 2-month extreme drought in 2021 reduced net ecosystem productivity by 444.0 ± 65.8 and 736.6 ± 57.8 grams of carbon per square meter at +9°C under aCO 2 and eCO 2 , respectively—228.6 ± 56.8% and 381.9 ± 83.4% of the reduction at +0°C under aCO 2 . This exacerbation was driven by warming-induced water table decline, prolonged low water tables, and CO 2 -enhanced substrate availability through increased plant carbon inputs. Findings indicate that future climate will greatly amplify carbon loss during extreme drought, reinforcing positive carbon-climate feedbacks.

Physics-guided machine learning (PGML) has become a prevalent approach in studying scientific systems due to its ability to integrate scientific theories for enhancing machine learning (ML) models. However, most PGML approaches are tailored to isolated and relatively simple tasks, which limits their applicability to complex systems involving multiple interacting processes and numerous influencing features. In this paper, we propose a Physics-Guided Foundation Model (PGFM) that combines pre-trained ML models and physics-based models and leverages their complementary strengths to improve the modeling of multiple coupled processes. To effectively conduct pre-training, we construct a simulated environmental system that encompasses a wide range of influencing features and various simulated variables generated by physics-based models. The model is pre-trained in this system to adaptively select important feature interactions guided by multi-task objectives. We then fine-tune the model for each specific task using true observations, while maintaining consistency with established physical theories, such as the principles of mass and energy conservation. We demonstrate the effectiveness of this methodology in modeling water temperature and dissolved oxygen dynamics in real-world lakes. The proposed PGFM is also broadly applicable to a range of scientific fields where physics-based models are being used.

Extreme drought events are predicted to increase with climate change, yet their impacts on ecosystem carbon dynamics under warming and elevated carbon dioxide (eCO2) remain unclear. In a peatland experiment with five warming treatments each under ambient carbon dioxide (aCO2) and eCO2 (+500 parts per million), a 2-month extreme drought in 2021 reduced net ecosystem productivity by 444.0 ± 65.8 and 736.6 ± 57.8 grams of carbon per square meter at +9°C under aCO2 and eCO2, respectively-228.6 ± 56.8% and 381.9 ± 83.4% of the reduction at +0°C under aCO2. This exacerbation was driven by warming-induced water table decline, prolonged low water tables, and CO2-enhanced substrate availability through increased plant carbon inputs. Findings indicate that future climate will greatly amplify carbon loss during extreme drought, reinforcing positive carbon-climate feedbacks.

Peatlands cover only 3% of Earth’s land surface but contain about 30% of the global soil carbon pool. The strong sensitivity of C cycle to environmental factors such as soil temperature and moisture has let to concerns about potential positive feedbacks to climate change. However, global models disagree as to the magnitude and spatial distribution of emissions, partially due to missing representations of peatland relevant processes and a scarcity of in situ observations. The Spruce and Peatland Responses Under Changing Environments (SPRUCE) experiment is a large‐scale climate change manipulation that focuses on the combined response of multiple levels of warming at both ambient and elevated CO2 concentration (eCO2), making it a valuable testbed for the broader modeling community to improve the diagnosis and attribution of C fluxes in peatland ecosystems. Currently, there are 11 models participating in the SPRUCE Model Intercomparison Project (SPRUCE-MIP). In the first stage, all model groups used observed ambient plot atmospheric forcing data to drive a model spin-up simulation with pre-industrial conditions, and a transient simulation with transient atmospheric CO2 concentrations and nitrogen deposition from 1850 to 2014. Then, measured plot-level meteorological forcing and CO2 concentrations from the 10 treatment enclosures and the ambient plot drove 11 transient simulations from 2015 to 2021. The total of 11 simulations represents five levels of temperature treatment with two CO2 levels, and ambient control plot with no enclosure. The five treatment temperatures are +0, 2.25, 4.5, 6.75, 9oC, and the two CO2 levels are ambient and +500 ppm. We evaluated the performance of multiple models against SPRUCE observations, such as the net ecosystem exchange (NEE) and CH4 fluxes warming responses under ambient and eCO2 conditions and found that there were wide spreads for warming responses among different models. We will further evaluate the model performances and quantify the associated uncertainties, which may have helpful implications for our understanding of the peatland C cycle and for future projections of Earth system models.

Abstract. Boreal peatlands store 13–32 % of the global soil carbon (C) stock, a service dependent on plant-mycorrhizal fungi associations. In these nutrient poor systems, ectomycorrhizal and ericoid mycorrhizal fungi supply up to >80 % of the nutrient requirements of their plant hosts, partly with mined nitrogen (N) and phosphorus (P) from soil organic matter that are otherwise inaccessible to plants. Despite the ecological significance, mycorrhizal associations are only represented in a few land surface or ecosystem models. We modify the peatland branch of version 2 of the Energy Exascale Earth System Land Model (ELMv2-SPRUCE) to replace the default photosynthesis-driven inorganic N and P (NP) uptake process with a more realistic representation of the process via three pathways: (1) direct inorganic NP uptake by uncolonized fine roots, (2) indirect inorganic NP acquisition and (3) indirect NP acquisition from organic sources by mycorrhizal roots. We systematically evaluated the performance of the default and modified models with field observations from a whole ecosystem warming and carbon dioxide fertilization experimental site: Spruce and Peatland Responses Under Changing Environment (SPRUCE), in northern Minnesota, USA. The modified model reduces the underestimation of the growth response of shrubs in the default model to warming from 40–80 % to 17–35 % and reduces the overall relative absolute error on C fluxes from 1.61 to 1.54. The improved growth response of shrubs to warming is accompanied by several-fold increase in direct inorganic NP uptake and decrease in fungal colonization rate. The modified model simulates a weaker transition of the ecosystem from C sink to C source under warming due to alleviation of plant nutrient limitation. Equifinality analysis shows the newly added parameters in the modified model can be constrained by the observed C fluxes. Sensitivity analysis shows the newly added parameters have stronger statistical interactions than the preexisting parameters in the default model. Overall, the modified model is an improvement over the default ELMv2-SPRUCE and will be a useful tool for understanding boreal peatland change.

The response of microbial communities that regulate belowground carbon turnover to climate change drivers in peatlands is poorly understood. Here, we leverage a whole ecosystem warming experiment to elucidate the key processes of terminal carbon decomposition and community responses to temperature rise. Our dataset of 697 metagenome-assembled genomes (MAGs) represents the microbial community from the surface (10 cm) to 2 m deep into the peat column, with only 3.7% of genomes overlapping with other well-studied peatlands. Community composition has yet to show a significant response to warming after 3 years, suggesting that metabolically diverse soil microbial communities are resistant to climate change. Surprisingly, abundant and active methanogens in the genus Candidatus Methanoflorens, Methanobacterium, and Methanoregula show the potential for both acetoclastic and hydrogenotrophic methanogenesis. Nonetheless, the predominant pathways for anaerobic carbon decomposition include sulfate/sulfite reduction, denitrification, and acetogenesis, rather than methanogenesis based on gene abundances. Multi-omics data suggest that organic matter cleavage provides terminal electron acceptors, which together with methanogen metabolic flexibility, may explain peat microbiome composition resistance to warming.

Climate change is reducing the amount, duration, and extent of snow across high-latitude ecosystems. But, in landscapes where persistent winter snow cover develops, experimental platforms to specifically investigate interactions between warming and changes in snowpack, and impacts on ecosystem processes, have been lacking. We leveraged a whole-ecosystem warming experiment in a boreal peatland forest to quantify how snow duration, depth, and fractional cover vary with warming of up to +9(degrees)C. We found that every snow-related quantity we examined declined precipitously as the amount of warming increased. The importance of deep, continuous snow cover for moderating shallow soil temperature is highlighted by an increase in soil temperature variance and the frequency of short-duration freeze-thaw cycles in the warmer plots. We used a paired-plot approach to estimate the magnitude of the snow-albedo feedback effect, and demonstrate that albedo-driven warming linked to reduced snow cover varies between December (+0.4(degrees)C increase in maximum air temperature) and March (+1.2(degrees)C increase) because of differences in insolation. Overall, results show that even modest future warming will have profound impacts on northern winters and cold-season ecosystem processes. Plot-level data from this warming experiment, and emergent relationships between warming and quantities related to snow cover and duration, could be of enormous value for testing and improving the representation of snow processes in simulation models, especially under future climate scenarios that are outside of the range of historically observed variability.

Accurate prediction of dissolved oxygen (DO) concentrations in lakes requires a comprehensive study of phenological patterns across ecosystems, highlighting the need for precise selection of interactions amongst external factors and internal physical-chemical-biological variables. This paper presents the Multi-population Cognitive Evolutionary Search (MCES), a novel evolutionary algorithm for complex feature interaction selection problems. MCES allows models within every population to evolve adaptively, selecting relevant feature interactions for different lake types and tasks. Evaluated on diverse lakes in the Midwestern USA, MCES not only consistently produces accurate predictions with few observed labels but also, through gene maps of models, reveals sophisticated phenological patterns of different lake types, embodying the innovative concept of "AI from nature, for nature".