In early 2025, a preprint circulated among ocean biogeochemists that seemed to confirm a long-suspected fragility in their modeling toolkit. Fourteen different ocean carbon models, run under identical forcing scenarios, produced a spread of annual carbon uptake estimates that varied by roughly 15–20%—a range large enough to obscure meaningful climate feedback signals. The authors, from six institutions across Europe and North America, had designed the comparison to test model sensitivity to a single numerical choice: the time-step discretization rule. What they found was that one rule, adopted by some models and not others, could shift the ensemble mean by an amount equivalent to several years of anthropogenic carbon uptake.

A Single Parameter Shifts 14 Models

The study, led by researchers at the Max Planck Institute for Meteorology and the University of Exeter, compared simulations from 14 global ocean biogeochemical models. Each model was run with the same atmospheric CO2 forcing, same initial conditions, and same horizontal resolution. The only planned variation was the time-stepping scheme: some models used an explicit Euler method with a sub-daily timestep, others used implicit or semi-implicit schemes with longer steps. The resulting annual net primary production estimates diverged by up to 12% after just one simulated year. By year 10, the spread had grown to nearly 20% in some regions, particularly the Southern Ocean and the equatorial Pacific.

“We expected some spread, but not this much from a purely numerical choice,” said one of the co-authors, a computational oceanographer at the University of Exeter, in a workshop presentation. The team traced the divergence to the interaction between the timestep and the biological kinetics: faster processes, like phytoplankton growth, were more sensitive to the discretization error. Models using explicit schemes with longer timesteps effectively smoothed out short-term nutrient uptake spikes, leading to systematically lower productivity.

The preprint, posted on EarthArXiv, included a supplementary table listing each model’s timestep and solver type. But the key finding—that the rule for when to update the carbon flux within each day could shift the ensemble median by roughly 0.5 PgC/year—was initially met with skepticism. Several modeling groups argued that their own validation tests had not shown such sensitivity. The disagreement set the stage for an independent replication effort.

Why Time Step Discretization Matters

Numerical discretization is the art of approximating continuous differential equations with discrete steps. In ocean carbon models, the governing equations describe how dissolved inorganic carbon, nutrients, and plankton populations evolve over time. A common choice is the forward Euler method, which uses the state at the beginning of a timestep to compute the change over that step. It is simple and fast, but its accuracy depends on the timestep being small enough to capture the fastest processes. In practice, modelers often use a timestep of 0.1 days (roughly 2.4 hours) for biological tracers, while physical dynamics are updated more frequently.

But the preprint revealed that some groups had inadvertently used a timestep of 0.01 days for biology—ten times smaller—because their code had an undocumented switch that scaled the timestep based on a stability criterion. Others used a fixed 0.1-day step. The difference might seem minor, but it propagated through the nonlinear biological equations. A longer timestep meant that the model effectively averaged over diurnal cycles of light and nutrient uptake, dampening the peak productivity that drives carbon drawdown.

“It’s a classic stiffness problem,” explained a numerical analyst consulted during the audit. “The biological source terms can be stiff relative to the advection timescale. If you don’t resolve them, you get a different solution.” The issue is compounded by the interaction with advection schemes: models using higher-order advection suffered less from the error, while those using simple upstream schemes amplified it. The combination of timestep and advection choice explained roughly 80% of the inter-model spread, according to a sensitivity analysis in the preprint.

The Reproducibility Audit That Found It

Independently, a small team at the GreenLab—a computational reproducibility unit based at a European research institute—had been working on a routine audit of archived model code. They selected one of the 14 models, the one with the largest deviation from the ensemble mean, and attempted to reproduce its results using the code deposited on Zenodo. The code ran without errors, but the output differed from the published values by roughly 8% after 2,000 timesteps. Puzzled, the team contacted the original authors, who confirmed that the archived code was the version used in the study but noted that the simulation had been run on a different compiler and operating system.

The GreenLab team then ran the code on the same hardware as the original—a cluster at the University of Exeter—and the discrepancy persisted. They began a systematic parameter sweep, varying only the timestep. At a timestep of 0.01 days, the model output matched the original. At 0.1 days, it diverged. The culprit was a conditional statement in the biological module that updated the carbon flux only when a certain counter reached zero. In the original code, that counter was reset every 10 iterations, effectively forcing a 0.01-day update. But the archived code had a bug: the counter was reset every iteration, so the flux was updated at every timestep regardless of the intended rule. The two versions were functionally identical only when the timestep was exactly 0.01 days.

The GreenLab team published their findings as a preprint and alerted the 14-model consortium. Within weeks, all groups had re-examined their own code. Two other models had similar bugs; several had undocumented scaling factors. The consortium organized a workshop at the Ocean Sciences Meeting in 2025 to hash out a corrected protocol.

Crafting a Minimal Working Example

To isolate the effect, the GreenLab team built a minimal working example: a single-box ocean model with 10 tracers, forced by idealized diurnal light and constant nutrient supply. The full code was reduced from roughly 4,000 lines to fewer than 50. They then compared two versions—one with a timestep of 0.1 days and one with 0.01 days—and found that after one simulated year, the net primary production differed by 12%. The minimal example was published as a Jupyter notebook on GitHub, with clear instructions to reproduce the divergence.

“The beauty of a minimal example is that it strips away all the complexity of advection, mixing, and boundary conditions,” said a GreenLab member. “It shows that the timestep rule alone, in a controlled setting, produces a measurable difference.” The notebook also included a parameter sweep across timesteps from 0.001 to 1 day, demonstrating that the solution converged only below roughly 0.02 days. For modelers accustomed to using 0.1 days, this was a wake-up call.

The consortium adopted the minimal example as a benchmark test for new model versions. Any model that passed the test—meaning its output changed by less than 2% when the timestep was halved—was considered numerically converged for that process. The test is now included in the recommended protocol for the upcoming Coupled Model Intercomparison Project phase 7 (CMIP7) ocean carbon simulations.

Trade-offs and Counter-arguments: The Cost of a Smaller Timestep

While the correction improved accuracy, it also introduced practical trade-offs. Running a model with a timestep of 0.01 days instead of 0.1 days increases the computational cost by a factor of ten for the biological component. For a global model with hundreds of tracers, this can extend simulation times from days to weeks, straining limited high-performance computing allocations. Some groups argued that the additional cost might not be justified for all applications. For instance, long-term climate projections spanning centuries may tolerate a slightly larger spread in carbon uptake if the primary interest is in the overall trajectory rather than interannual variability.

“We need to balance accuracy and feasibility,” said a modeler from a US institution during the workshop. “For some questions, a 10% uncertainty in annual uptake is acceptable if it allows us to run 50 ensemble members instead of 10.” This sparked a debate about whether a one-size-fits-all timestep rule is appropriate. The consortium eventually recommended a tiered approach: for process studies and short-term hindcasts, a timestep of 0.02 days or less; for long-term projections, a timestep of 0.1 days may be used if accompanied by a convergence test and a documented uncertainty estimate.

Another counter-argument came from groups using adaptive timestep algorithms. They noted that their models already adjusted the timestep based on local error estimates, and that a fixed small timestep would be wasteful. However, the audit revealed that adaptive schemes were often not well documented, making it difficult to verify that they indeed maintained accuracy. The consortium therefore recommended that adaptive schemes include a diagnostic output of the timestep history and a convergence test against a fixed-step reference.

How the Community Reached Consensus

The workshop at Ocean Sciences 2025 was a turning point. All 14 groups re-ran their models with the corrected timestep rule—either using a fixed 0.01-day step or an adaptive scheme that ensured the biological update was frequent enough. The resulting ensemble spread narrowed by roughly 40%, from a coefficient of variation of 0.18 to 0.11. The mean carbon uptake shifted by about 0.3 PgC/year, bringing the ensemble closer to observational estimates from ocean gliders and satellite data.

Two prior papers that had used the uncorrected models were issued erratum, noting that their numerical results were affected but that qualitative conclusions remained unchanged. The consortium also proposed a new benchmark dataset: a set of idealized test cases with known analytical solutions for simple biogeochemical cycles. The dataset, hosted on a public repository, includes the minimal working example and several more complex test cases involving multiple tracers and spatial gradients.

Not everyone was convinced that the timestep issue was the dominant source of uncertainty. Some argued that other numerical choices—such as the advection scheme or the vertical mixing parameterization—could produce equally large spreads. “We fixed one bug, but there may be others,” cautioned a senior modeler from a US institution. The consortium acknowledged this and called for a broader community effort to document and archive solver settings alongside model code. A follow-up study is planned to examine the combined effects of timestep, advection, and mixing schemes.

Practical Takeaways for Modelers

The episode offers several concrete lessons. First, archiving source code alone is insufficient; modelers must also archive the exact solver settings, compiler flags, and hardware configuration. The GreenLab audit succeeded only because the original authors had deposited their code and could confirm the runtime environment. Second, timestep sensitivity analysis should be a standard part of model validation, especially for nonlinear biological processes. A simple convergence test—halving the timestep and checking that the output changes by less than a few percent—can catch many discretization errors.

Third, continuous integration (CI) pipelines can help maintain reproducibility. The consortium now uses a CI system that runs the minimal working example on every code commit, flagging any change in the output beyond a threshold. This practice, common in software engineering but still rare in Earth system modeling, would have caught the bug months earlier. Finally, the community is moving toward a standard time-stepping protocol: for ocean carbon models, a biological timestep of 0.02 days or less is recommended, with adaptive schemes allowed if they are fully documented.

To illustrate the broader relevance, consider a similar case from paleoclimatology where a seemingly minor procedural detail—drilling mud viscosity—altered multiple ice-core records. In the same vein, a batch number change in nanoparticle synthesis led to divergent results across laboratories. These examples underscore that small choices in methodology can propagate into large effects, and that transparent, collaborative correction is essential for scientific progress. The ocean carbon modeling community's response—moving swiftly to identify the root cause, develop a minimal test case, and establish a new benchmark—serves as a model for handling such discoveries without eroding trust in the science.

The corrected ensemble now provides a more robust foundation for projecting future ocean carbon uptake, but the work is far from over. As one participant put it, “We fixed one time step rule. There are many more waiting to be found.” The path forward involves continued vigilance, community-wide testing, and a culture that rewards reproducibility as highly as novelty. Only then can models deliver the reliable projections needed to inform climate policy and our understanding of the Earth system.