The research team, led by Sagar Gautam of Sandia National Laboratory’s Bioscience Division together with Umakant Mishra, and including soil scientists Rattan Lal and Klaus Lorenz of The Ohio State University, Jinyun Tang of Lawrence Berkeley National Laboratory, DeAnn Ricks Presley of Kansas State University, and Alan J. Franzluebbers of the USDA Agricultural Research Service, simulated a conventional corn–soybean rotation across cultivated lands of the continental United States at a 4 square kilometer spatial resolution. Their ensemble integrated three models with fundamentally different architectures: DAYCENT, a daily time-step descendant of the classic CENTURY model; DNDC, which focuses on coupled carbon and nitrogen biogeochemistry; and ECOSYS, a highly mechanistic model that resolves plant physiology, root growth, soil physics, and microbial dynamics in detail. Each model was fed the same harmonized environmental inputs, including daily weather from the DAYMET dataset, soil properties from the Soil Survey Geographic Database, and land cover information from the National Land Cover Database.

The choice of models was deliberate. DAYCENT represents soil organic matter as a set of pools with different turnover rates, driven by decomposition rates that depend on soil temperature, moisture, and texture. DNDC simulates the full crop life cycle and the exchange of carbon dioxide, methane, and nitrous oxide between soil and atmosphere. ECOSYS goes further, explicitly modeling photosynthesis, plant hydraulics, nutrient exchange between microbes and plants, and the physical protection of carbon within soil aggregates. Because these frameworks interpret the same field conditions in different ways, their predictions diverge, and that divergence is precisely what an ensemble can exploit. By averaging across structurally different models, the ensemble mean captures the central tendency of plausible system responses while the spread among models quantifies structural uncertainty, a measure of how much conclusions depend on which assumptions a model makes.

To test the framework, the researchers calibrated each model against two independent observational datasets: county-level crop yields from the US Department of Agriculture’s National Agricultural Statistics Service and soil organic carbon measurements from the Rapid Carbon Assessment, a national database of field-sampled soil carbon. Notably, the team deliberately excluded performance metrics such as the Nash–Sutcliffe Efficiency and the coefficient of determination, which can be misleadingly inflated in large regional comparisons where predictions track a strong regional mean. Instead, they relied on root mean square error and its normalized form, and they ran a stratified cross-validation in which half the data were used for calibration and half held out for validation. The error distributions were essentially identical between the two subsets, with ensemble soil carbon errors of 4.7 versus 4.6 kilograms of carbon per square meter and yield errors matching to within a tenth of a megagram per hectare, demonstrating that the framework was robust against overfitting.

The results were striking. For corn, the ensemble median yield of 7.7 megagrams per hectare closely matched the observed median of 8.5 megagrams per hectare, and the ensemble achieved the lowest error of any approach, with a root mean square error of 3.0 megagrams per hectare. Individual models told a more complicated story: DAYCENT and DNDC performed moderately well with errors of 3.8 and 5.0 megagrams per hectare respectively, while ECOSYS systematically underpredicted corn yields, producing a median of only 4.4 megagrams per hectare. For soybean, the ensemble again led the field with a root mean square error of 1.0 megagram per hectare and a median of 2.5 megagrams per hectare against an observed median of 3.0. The pattern repeated for soil carbon: the ensemble produced a median stock of 5.5 kilograms of carbon per square meter in the top 30 centimeters of soil with the lowest error of 2.5 kilograms, compared with 4.8 kilograms measured in the field, while DNDC overestimated and ECOSYS underestimated the same quantity.

Why did the individual models disagree so much? The study’s analysis reveals how deeply model structure shapes outcomes. ECOSYS’s underprediction of yields likely stems from its detailed rhizosphere sub-model, in which nitrogen uptake by roots is governed by ion diffusion and mass flow. When the researchers applied standardized nitrogen input data across the continent, this mechanistic sensitivity triggered localized nitrogen stress more extreme than observations support. The authors also note that using a single, uniform set of crop phenology parameters across the entire country penalizes ECOSYS, whose simulated crop development is highly sensitive to thermal time and maturity group, both of which vary regionally. DNDC’s tendency to overestimate soil carbon in some regions and project losses in others reflects its Double Monod kinetic framework, in which abundant mineral nitrogen can accelerate the decomposition of existing organic matter to satisfy microbial stoichiometric demands faster than new crop residues can be converted into stable humus.

DAYCENT, by contrast, takes a more conservative pool-based approach in which nitrogen primarily boosts plant productivity and residue inputs rather than accelerating the decay of stabilized carbon pools, which helps explain its relatively strong performance for soil carbon in the Midwest. ECOSYS assumes that high carbon inputs are largely stabilized as microbial biomass and physically protected within soil aggregates rather than respired away. These are not trivial technical differences; they represent genuinely different scientific hypotheses about what controls whether carbon entering the soil stays there. The ensemble framework does not resolve that debate, but it makes the disagreement explicit, and the ensemble mean benefits from error compensation, with each model’s biases partially canceling those of the others.

The spatial maps produced by the ensemble carry practical implications. Across the 2014 to 2023 simulation period, all three models captured the broad pattern of high corn and soybean productivity in the Midwest Corn Belt, particularly Iowa, Illinois, and Indiana, but diverged elsewhere, with DNDC projecting relatively higher corn yields in the Mississippi River Basin and the Southeast while ECOSYS favored the Northern Plains. The ensemble smoothed these model-specific extremes into a coherent national picture. For soil carbon, the ensemble projected gains in the Midwest and Southeastern regions and losses in the Great Plains and the Western United States, with Corn Belt changes ranging roughly from minus 2 to plus 2 tons of carbon dioxide equivalent per hectare per year and a modest net positive average. That spatial heterogeneity, the authors argue, underscores the need for region-specific management rather than one-size-fits-all carbon policies.

The researchers are careful about the limits of their work. The 4 kilometer resolution, chosen to balance computational cost against the need to capture continental-scale variability, cannot resolve field-level heterogeneity in soils, management, and microclimate, which means the framework is not yet ready for farm-scale carbon credit quantification. The three-model ensemble is also small compared with large intercomparison initiatives such as the Agricultural Model Intercomparison and Improvement Project, so the full range of structural uncertainty is probably wider than what these simulations capture. The authors describe their results as an upper bound of current process-based modeling capability at this scale, and they caution that applications to specific regenerative agriculture programs remain preliminary until models are refined and higher-resolution data become available.

Even with those caveats, the study marks a significant step toward trustworthy carbon accounting at national scale. Soil carbon measurement, reporting, and verification underpins emerging carbon farming programs, and models like DAYCENT and DNDC are already used in national greenhouse gas inventories and carbon credit calculations. By demonstrating that a multi-model ensemble reduces bias, lowers error, and communicates uncertainty transparently, the researchers have provided a scalable, data-driven template that could be extended to other crops, other practices, and other countries. As climate variability intensifies and demand grows for verifiable soil carbon sequestration, the message of this work is clear: no single model should be trusted alone, and the path to reliable predictions runs through the honest integration of many.

Subject of Research: Multi-model ensemble simulation of crop yields and soil organic carbon dynamics across US croplands

Article Title: Ensemble agroecosystem modeling enhances predictions of crop yields and soil carbon across the United States

Article References: Gautam, S., Jung, C. G., Lal, R., Lorenz, K., Tang, J., Presley, D. R., Franzluebbers, A. J., & Mishra, U. (2026). Ensemble agroecosystem modeling enhances predictions of crop yields and soil carbon across the United States. SOIL, 12(2), 821-833. https://doi.org/10.5194/soil-12-821-2026

Image Credits: AI Generated

DOI: 10.5194/soil-12-821-2026

Keywords: agroecosystem modeling, soil organic carbon, crop yields, multi-model ensemble, DAYCENT, DNDC, ECOSYS, carbon sequestration, corn-soybean rotation, carbon accounting, United States croplands, uncertainty quantification