How Do You Model Soil Carbon Change Over Time?

A soil carbon model takes a description of a system (climate, soil, management) and returns a trajectory: how the carbon stock is expected to evolve under that description. That trajectory is how organisations justify capital allocation to regenerative programmes, forecast Scope 3 emissions and removals under the GHG Protocol Land Sector and Removals Guidance, and, where the project structure warrants it, file claims on voluntary carbon markets.

The question “which model should we use?” has no universal answer. The right question is narrower: which model is fit for the decision, on the system at hand, given the data the project can realistically get? Answering it is the single most important judgment call in a soil carbon modelling engagement.

Underneath the acronyms, the workhorse models share one mechanism: soil carbon sorted into pools that each decay at their own rate, fed by a stream of fresh plant and manure inputs. Understand that mechanism and most of the practical questions answer themselves, including the two that quietly decide a decadal forecast, namely where you assume the pools start and how wide the honest answer is. The three tools in this piece run a compact version of that mechanism, so the behaviour is something you can move rather than only read about.

What modelling actually delivers, and what it doesn't

A soil carbon model is a structured hypothesis about how carbon moves through the system. Its outputs are conditional: given these inputs, here is the expected trajectory and its uncertainty. A model does not measure the soil; it predicts it.

This matters because much of the disagreement between project developers and verifiers comes from confusing the two1. A modelled trajectory substituted for field data in annual reporting is a compliance risk. A modelled trajectory used alongside field data, with uncertainty ranges explicit, is standard practice under the GHG Protocol Land Sector and Removals Guidance and under most voluntary market methodologies.

How these models actually work

Soil carbon is not one substance decaying at one rate. A fresh root and a centuries-old humic complex are both “soil organic carbon,” yet one turns over in months and the other outlives the person measuring it. Every model in common use handles this the same way. It sorts the carbon into a handful of pools, each with its own turnover time, and tracks them separately2.

Within each pool, decomposition follows first-order kinetics: the amount lost in a given month is proportional to the amount present, so a pool left to itself decays along an exponential curve set by a rate constant k.

The rate constant k is a property of the pool: fast for fresh residues, slow for humified matter, effectively zero for the inert fraction. The function f scales k up or down with the conditions the soil is actually in. Warmth speeds decomposition, moisture stress slows it, and a living canopy retards it relative to bare ground. That is why the same residue input builds carbon under a cover crop in a cool temperate season and loses it in a hot, dry one.

Not all of the decomposed carbon leaves the soil. A fraction is respired to the atmosphere as CO₂; the rest is stabilized into microbial biomass and humified compounds that feed back into the slow pools. Clay content governs that split, because finer-textured soils protect more of the processed carbon. This is why a clay fraction is a required input in every model below3.

Put those pieces together and a soil under steady management drifts toward a steady state, the stock at which annual inputs exactly balance annual losses. Raise the inputs and the steady state rises; warm the climate or lower the clay and it falls. The approach is slow, paced by the turnover time of the slowest active pool, which is why credible soil carbon gains are quoted over decades rather than years. The tool below runs this mechanism forward. Move the input, clay, and temperature and watch where the stock settles and how long it takes to get there.

Two features of that curve matter for a project. The destination is fixed by the inputs and the site, not by where the soil happens to start, so two fields brought under the same management converge on the same steady state from different directions. And the journey is long: even a large change needs decades to express itself, and the first years move slowly. A modelled trajectory that reaches a new equilibrium in five years is describing an artefact of its setup, not a soil.

The four candidate models

Four frameworks dominate practical work. Each was designed for a specific class of question, and each carries the fingerprint of the system it was built on. RothC grew out of the Rothamsted long-term experiments on temperate arable topsoil2. AMG was first calibrated on French cropping trials and rebuilt on more than forty European long-term sites in its current form45. Century was designed to explain organic matter levels in Great Plains grasslands, with nitrogen, phosphorus, and sulphur coupled to the carbon cycle6, and DayCent is its daily-time-step sibling, built to resolve the trace-gas fluxes a monthly model cannot7. Using one outside its home domain is possible but expensive in credibility. On testing whether a new site is inside that domain at all, see Representative of What?

All four have been benchmarked against long-term experimental data across dozens of sites. The classic comparison ran nine soil organic matter models against seven long-term datasets and found no single winner: performance depended on the site and the question8. More recent ensemble work reaches the same verdict and adds a warning: the spread between models is itself a source of uncertainty that a single-model projection hides11. RothC and Century have the longest international track record; AMG is the strongest option for French and other European cropping systems; DayCent is the specialist tool when N₂O emissions are in scope.

Choosing the right model for the question

Four practical questions determine the choice.

Is nitrogen in scope?

If the decision depends on N₂O emissions, fertilizer-use efficiency, or the interaction between nitrogen management and carbon dynamics, RothC is structurally the wrong choice; it has no N pool. Century or DayCent are the honest answers. If the decision is purely about soil carbon under stable nitrogen management, RothC is defensible and simpler.

Is the system water-limited?

RothC's default water modifier underperforms in dryland and Mediterranean systems, where biological activity is limited by moisture for much of the year9. A modified RothC calibration (Farina et al.) exists and is the practical choice in those contexts. Century and DayCent handle water limitation natively through their submodels.

What time resolution is needed?

Monthly is sufficient for a 20-year decadal trajectory. Daily is required when the outcome of interest is a rapid flux: trace-gas emissions after a fertilizer application, for instance. RothC, AMG, and Century operate monthly; DayCent operates daily.

What data can the project realistically produce?

A model that needs inputs the project cannot supply is a model that will run on assumptions. If daily weather is not available, DayCent will operate on a downscaling, which is defensible if the downscaling is disclosed. If residue inputs and management history are unknown, every model is in the same position: the uncertainty from missing inputs dominates the uncertainty from the model.

Calibration, validation, and knowing when a model fits

A model that has never been tested against measured carbon on a comparable system is a hypothesis, not evidence. Calibration tunes a model's parameters to reproduce a known dataset; validation asks the harder question of whether it then predicts data it never saw. The two are routinely confused, and a model that fits its calibration data beautifully can still fail on an independent site.

The instruments for this are the long-term experiments, some running for more than a century, where soil carbon has been tracked under fixed treatments. Model skill is reported against them with familiar statistics: root mean square error for the size of the miss, and model efficiency for whether the model beats simply predicting the mean. The nine-model comparison that set the standard for this work found agreement on some sites and sharp divergence on others, with no framework dominant everywhere8. A 26-model ensemble on bare-fallow soils reached a sharper conclusion still: over long projections the choice of model can matter as much as the choice of management11.

Two practical rules follow. Validate on time-series data from the region and system you are modelling, not on the global average, because a model calibrated on temperate arable soils carries no guarantee on tropical agroforestry. And treat the applicability domain as a real boundary: a model asked to extrapolate beyond the conditions it was tested on returns a number with false confidence, which is the subject of Representative of What? When a project needs model-based accounting at the level national inventories use, the IPCC calls this Tier 3 and asks for exactly this evidence, a model validated against measurements representative of the system12. The reviews that track the field's direction make the same point from the research side10.

Initialization: the silent source of error

Every soil carbon model partitions the stock into pools with different turnover rates: fast, slow, and near-inert. The initial split across pools is rarely measurable and is usually estimated. For RothC, the Falloon equation estimates the inert pool from total SOC; it is a statistical approximation with a material effect on 30-year projections13.

The equation above is a statistical fit, not a physical measurement. Its parameters carry uncertainty that compounds over decadal projections. The more defensible alternative is spin-up: run the model to equilibrium under a reconstructed historical management regime, then use the resulting pool distribution as the starting point. The reconstruction is the hard part, and is also the step most commonly glossed over in project documentation. A 50-year historical back-cast under a realistic rotation reduces the effect of initialization on forecast trajectories by an order of magnitude compared with default-pool initialization14.

The tool below shows why this is not a technicality. Fix a measured stock, fix the future management, and change only how the inert pool is assumed. The inert fraction never gains or loses carbon, so putting more of the stock there leaves less that can respond. The same measurement, split two ways, can produce a projected gain under one assumption and a projected loss under the other. That is the whole decision, decided by a number nobody measured.

Reading outputs as envelopes, not lines

A soil carbon trajectory with a tight 95 percent confidence interval either reflects a well-calibrated, data-rich system, or underreporting. Monte Carlo propagation of input uncertainty (climate variability, clay content, residue inputs, initialization) produces a trajectory envelope that is usually substantially wider than the mean line. That envelope is the honest deliverable. Strip it out and the remaining number is, at best, a plausible central estimate.

For scenario comparison work (baseline vs. intervention, for instance), the useful output is the difference between scenarios, with its own propagated uncertainty. Much of the input noise is shared between scenarios and cancels in the difference, which is often tighter than either individual trajectory. This is the right framing for nearly all carbon project feasibility questions.

The tool below makes this concrete. Toggle between showing the two scenarios separately, each with an envelope wide enough to swallow the signal, and showing their difference, where the shared climate and soil noise cancels and a defensible number emerges. The more the baseline and intervention share the same conditions, the tighter that difference becomes. This is the statistical reason serious MRV frameworks reward paired designs and control sites rather than absolute before-and-after claims15.

The frontier: what the pool models leave out

The four workhorse models all descend from the same idea, conceptual pools defined by turnover rate. It is a powerful abstraction, but the pools are not things you can go out and measure. You cannot sieve a soil into its decomposable and humified fractions; they are bookkeeping compartments inferred from behaviour. For twenty years that gap has driven a second generation of models built on quantities you can measure.

The shift in the underlying science is from chemistry to biology and mineralogy. Persistence, on the current view, is less about a molecule being intrinsically hard to break down and more about whether microbes can reach it and whether mineral surfaces hold it18. Microbially explicit models such as MIMICS and Millennial put microbial biomass and mineral-associated carbon at the centre, and replace inferred pools with fractions that can, in principle, be measured in a lab1617. They respond differently from first-order models under warming, and the difference is not small.

None of this retires RothC or Century for project work. The newer models are harder to parameterise, less benchmarked against long-term field data, and not yet written into MRV standards. But they mark where the uncertainty in a decadal projection actually lives, and a practitioner who treats the pool structure as physical truth rather than a useful fiction will be caught out by the next decade of the literature.

From trajectory to defensible claim

A trajectory only becomes an asset when a standard will accept it, and the frameworks that matter here converge on one posture: model to plan and to interpolate between measurements, but measure to claim. The GHG Protocol Land Sector and Removals Guidance treats modelled removals as reportable only when paired with a measurement and monitoring programme and with uncertainty stated1. National inventories using model-based Tier 3 accounting carry the same requirement of validation against representative data12. The reference paper on soil-carbon MRV is explicit that models reduce sampling cost but do not remove the need to measure15.

So the honest deliverable of a modelling engagement is rarely a single number. It is a trajectory with its envelope, a clear statement of which model was used and why, a documented initialization, and a plan for the measurements that will confirm or correct it. Continental-scale studies model millions of hectares on exactly this logic19, and it holds at field scale too: the model carries the projection between the points where the soil is actually sampled.

Key takeaways