The first key motivation for C4MIP is to document the changes in magnitude of the feedback parameters that characterize the response of the carbon cycle and their spread across models through time. In this respect, C4MIP aims to calculate the magnitude of the carbon-concentration and carbon–climate feedbacks in a manner similar to Friedlingstein et al. (2006) or Arora et al. (2013) and as discussed in Sect. 3.1 using results from the idealized 1 % per year increasing CO2 experiments.

The 1pctCO2 experiment has gained recognition as a standard CMIP simulation and it is one of the DECK simulations for CMIP6 (Eyring et al., 2016a). The 1pctCO2 experiment is now routinely used to characterize the transient climate response (TCR) defined as the change in globally averaged near-surface air temperature at the time of CO2 doubling as well as the transient climate response to cumulative emissions (TCRE) defined as change in globally averaged near-surface air temperature per unit cumulative CO2 emissions at the time of CO2 doubling (Gillett et al., 2013). In addition, since the 1pctCO2 simulation does not include the confounding effects of changes in land use, non-CO2 greenhouse gases, and aerosols it provides a clean controlled experiment with which to compare carbon–climate interactions across models. Its backwards compatibility enables direct comparison of models with previous generations, which has been hindered previously as the scenario-dependence of the feedback metrics has prevented a like-for-like comparison (Gregory et al., 2009).

C4MIP will use partially coupled simulations to isolate and quantify the sensitivity of carbon cycle components to climate and CO2 separately and also the potentially large non-linear combination of these two components (Gregory et al., 2009; Schwinger et al., 2014). Simulations with only the carbon cycle model components experiencing rising CO2 (biogeochemically (BGC) coupled) and the radiation model components experiencing rising CO2 (radiatively (RAD) coupled) are used to quantify the carbon-concentration and carbon–climate feedbacks. Spatial patterns of these metrics can also be calculated (e.g. Roy et al., 2011; or Fig. 6.22 of the last IPCC WG1 assessment report Ciais et al., 2013) to establish areas of model agreement or disagreement.

The historical simulations will be used for evaluation of the components of the carbon cycle (ocean and terrestrial carbon fluxes, anthropogenic carbon storage in the ocean, atmospheric CO2 growth rate and variability). ESMs have increased rapidly in complexity but evaluation has not kept pace. Some evaluation of the carbon cycle was already performed in CMIP5 (e.g. Anav et al., 2013; Bopp et al., 2013; Hoffman et al., 2014), highlighting significant biases in key quantities in many ESMs. There is increasing need to develop evaluation techniques and activities, applied consistently and routinely across models, at both fine scales (process-level, “bottom-up” evaluation) and large scales (system-level, “top-down” evaluation”), as well as using complementary data streams relating to (bio)physical and biogeochemical processes to evaluate the ensemble of simulated processes (e.g. Luo et al., 2012; Foley et al., 2013).

Evaluation of ocean carbon cycle components of ESMs has been classically based on the use of the monthly surface pCO2 climatology of Takahashi et al. (2009), derived from more than 3 million in situ ocean pCO2 measurements, as in Pilcher et al. (2015) for an evaluation of pCO2 seasonality of the CMIP5 ESMs. This evaluation is complemented by the use of additional climatological gridded products, as in Anav et al. (2013), with model–data comparison for related physical variables (e.g. mixed layer depth) or biological variables (e.g. net primary production). In the past few years, ESM evaluation has extended in many directions, making use of advanced observation-based gridded products (e.g. the three-dimensional distribution of anthropogenic carbon in the ocean from Khatiwala et al., 2013) and ocean databases with millions of in situ point measurements (e.g. with the Surface Ocean CO2 Atlas (SOCAT) as in Tjiputra et al. (2014) for CMIP5 ESMs), or developing new techniques for model–data comparisons (e.g. water-mass framework; Iudicone et al., 2011).

In the coming years, the increasing complexity of marine biogeochemical schemes used in ESMs will call for more advanced model–data comparison strategies. These will include the use of new data sets, such as biomass data for plankton functional types (MAREDAT; Buitenhuis et al., 2013) or ocean distribution of the micro-nutrient iron (Tagliabue et al., 2012).

Evaluations of land surface components of ESMs have often used gridded flux products (e.g. Bonan et al., 2011; Anav et al., 2013; Piao et al., 2013) obtained by extrapolating the FLUXNET measurement network of biosphere–atmosphere exchanges (e.g. Jung et al., 2011), for instance to constrain modelled spatial and seasonal distribution of gross primary production (GPP). Such products are convenient for such model evaluations because those are available at a resolution comparable to that of the models and because they retain the pertinent patterns of the observed fluxes while abstracting from measurement noise, local site representativeness and other possible site-specific features. Yet it is important to bear in mind the limitations of the “upscaled” flux and stock products and to tailor the model evaluation to robust patterns that the individual products are ideally suited for. Insights may also be gained from evaluation of functional patterns and sensitivities to certain climate forcing variables. For example the spatial sensitivity of GPP with mean annual precipitation in the water-limited domain, and the temperature sensitivity of ecosystem respiration (Mahecha et al., 2010).

While data-model comparisons of fluxes are important, they alone cannot constrain longer-term dynamics and associated climate–carbon cycle feedbacks. In addition, consideration of residence times is crucial, which together with carbon fluxes jointly determine the stores. Analysis of CMIP5 ESMs revealed unacceptably large errors in land carbon stores (both in living biomass and soil organic matter) (Anav et al., 2013). Future simulation results were found to depend on the initial conditions as well as the model sensitivity to changes (Todd-Brown et al., 2014) and therefore better evaluation and constraint of carbon stores is seen as vital. Xia et al. (2013) showed the importance of residence time in determining carbon stores and Carvalhais et al. (2014) showed the mismatch between CMIP5 ESMs and an observationally derived data set of land-carbon residence times. As more observations become available (Saatchi et al., 2011; Baccini et al., 2012; Avitabile et al., 2015; FAO, 2012; Batjes et al., 2012; Hengl et al., 2014) as well as data constrained products such as residence time (Bloom et al., 2016), we stress the importance of rapid development and application of evaluation techniques to ESMs.

Carbon isotopes (carbon-13 and carbon-14) provide unique insights into the mechanisms and timescales of carbon cycling. Differences between the isotopic fractionation of carbon from dissolution in the ocean and from photosynthetic assimilation on land have enabled atmospheric observations of the 13C / 12C ratio (δ13C) in atmospheric CO2 to be used in differentiating land and ocean carbon fluxes (Ciais et al., 1995; Joos et al., 1998; Rubino et al., 2013). The perturbation of the 14C / C ratio (Δ14C) in atmospheric CO2 from nuclear weapons testing in the 1950s and 60s has provided a valuable tracer of carbon turnover rates in terrestrial carbon pools (Trumbore, 2000; Naegler and Levin, 2009), and the rates of air–sea exchange and ocean mixing, including constraints on ocean CO2 uptake (Matsumoto et al., 2004; Sweeney et al., 2007; Graven et al., 2012). Integration of carbon isotopes into ESMs is an emerging activity and we request the reporting of carbon isotopic variables for the first time in C4MIP. Carbon isotopes are also included in OMIP (Orr et al., 2016). ESMs that simulate carbon isotopes are requested to report fluxes and stocks of carbon isotopes in their land and ocean components. This will enable comparison between models currently simulating carbon isotopes and their evaluation by observations, as well as encourage future development of carbon isotopes in ESMs. Simulation of carbon isotopes in C4MIP is expected to provide novel insights on ocean mixing and air–sea exchange, marine ecosystem change, plant water use efficiency and stomatal closure especially during drought periods, and terrestrial carbon residence times.

Historical simulations will be needed to explore potential emergent constraints from observations on the future response of the carbon cycle, with a particular focus on carbon cycle feedbacks. Recent studies showed the potential of observed interannual CO2 variability to constrain the future tropical land carbon cycle sensitivity to climate change (Cox et al., 2013; Wenzel et al., 2014).

In the same way that Earth system modelling has become an internationally collaborative activity involving shared expertise and development of tools, we also expect that evaluation techniques will evolve in this way. Community evaluation activities such as ILAMB (http://www.ilamb.org/) and ESMValTool (Eyring et al., 2016b) look likely to become increasingly useful for addressing the complexities of multi-model ESM evaluation.

While idealized experiments are useful for intercomparison of climate–carbon interactions across multiple models, they do not take into account the effect of non-CO2 GHGs, aerosols, and land-use change, all of which affect the behaviour of the carbon cycle in the real world. In contrast, the scenarios considered by the ScenarioMIP are internally coherent in all aspects of anthropogenic forcings. Within each socio-economic storyline, changes in fossil fuel CO2 emissions are consistent with those in aerosols emissions, N deposition, and changes in land-use areas, all of which are based on plausible assumptions of demographic and economic development in the future. This plausibility is of special interest to policymakers. Scenarios also indicate the range of possible future developments and opportunities for mitigation and adaptation; this information is used widely in climate impact analyses.

The scenario simulations, therefore, provide more realistic conditions compared to the idealized 1 % experiments due to their plausibility of anthropogenic forcings as well as the longer timescale over which the CO2 increase occurs. Since shared socio-economic pathway (SSP) scenarios include all forcings, their climate and biogeochemical effects are able to influence the atmosphere–surface carbon exchange for both land and ocean components. Emission-driven historical and the future SSP5-8.5 simulations replicate a more realistic model setting where ESMs are directly forced by anthropogenic CO2 emissions, allowing for the carbon cycle feedbacks to impact on atmospheric CO2 and simulated climate change. These will be compared with the concentration-driven equivalents in ScenarioMIP and additionally will form a baseline control experiment for analysis of alternative future land-use scenarios in LUMIP (Lawrence et al., 2016).

The proposed biogeochemically coupled versions of the historical and future SSP5-8.5 in Sect. 3.1, in which CO2 induced warming is not accounted for, when compared to their fully coupled versions will allow us to investigate the effect of CO2 induced warming on atmosphere–land and atmosphere–ocean CO2 fluxes over the 20th and 21st century and beyond (Randerson et al., 2015). An important objective with these simulations will be to identify how land and ocean contributions to feedbacks and compatible emissions evolve century by century from sustained increases in ocean heat content and thawing of permafrost soils.

ScenarioMIP (O'Neill et al., 2016) acknowledges scientific and policy interest in a scenario with a substantial overshoot in radiative forcing during the 21st century. As such they include a tier-2 concentration-driven scenario called SSP5-3.4-OS: an overshoot pathway, which follows SSP5-8.5 up to 2040, followed by aggressive mitigation to reduce emissions to zero by about 2070, and by substantial negative global emissions thereafter. The carbon cycle response to peak-and-decline CO2 levels is likely to differ from the response to continued strong increases in CO2. The 21st century airborne fraction from CMIP5 models varied substantially between RCPs, with RCP2.6 in particular having a much lower airborne fraction than the 20th century or other RCPs (Jones et al., 2013). However, to date there have been no coordinated experiments to quantify the carbon-cycle feedback components in such a scenario. Hence, for C4MIP we include a BGC simulation of the SSP5-3.4-OS scenario.

To participate in C4MIP a climate model must have the capability to run with an interactive carbon cycle. This means it must simulate both terrestrial and marine carbon cycle processes, and it must simulate the exchange of CO2 between the land/ocean and the atmosphere in order to prognostically simulate the evolution of atmospheric CO2. Some C4MIP simulations prescribe a concentration of CO2 in the atmosphere as a boundary condition and simulate the changes in carbon fluxes and stores in response. Other simulations prescribe emissions of CO2 to the atmosphere (from human activity) as an external forcing and require the model to also simulate the evolution of atmospheric CO2. A model must be able to run in both these configurations in order to perform the C4MIP simulations. The evolution of atmospheric CO2 concentration can be simulated by assuming that CO2 is completely well mixed with the same globally averaged concentration everywhere in space or by transporting CO2 as a three-dimensional tracer. This choice is up to the modelling groups. Throughout this document we refer to the former – prescribing atmospheric CO2 concentration as a boundary condition – as a “concentration-driven” simulation, and the latter – prescribing emissions and in turn simulating the CO2 concentration – as an “emissions driven” simulation. IPCC AR5 WG1 Ch.6 Box 6.4 described the use of these configurations in some detail (Ciais et al., 2013). Figure 6.4 from that Box is reproduced here for reference (Fig. 2). Although the same terminology (concentration-driven or emissions-driven) can be applied to aerosols or non-CO2 GHGs this paper focuses only on CO2.

Before beginning the simulations described below, a model must be spun-up to eliminate any long-term drift in carbon stores or fluxes. Indeed, it has been shown recently that the large diversity in spin-up protocols used for marine biogeochemistry in CMIP5 ESMs contribute to large model-to-model differences in simulated fields, and that drifts have potential implications on model performance assessments in addition to possibly aliasing estimates of climate change impacts (Séférian et al., 2016). Separate spin-up simulations should be performed for both concentration-driven and emission-driven configurations. There are many possible techniques to ensure that a model's carbon fluxes and pools exhibit minimal drift. These include simply performing very long simulations, running components offline from the coupled system, numerical acceleration techniques or semi-analytical schemes such as described by Xia et al. (2012). The choice of technique is up to the modelling groups and there is no requirement to submit data from the spin-up period, but a proper documentation of the spin-up technique and duration is required. The test of whether a model is spun-up properly and exhibits minimal drift will be based on the performance of the piControl simulation. It is suggested that the model first be spun-up in concentration-driven mode and this state can be used as an initial basis for the emission-driven spin-up.

Our definition of an acceptably small drift in a properly spun-up model is that land, ocean, and atmosphere carbon stores each vary by less than 10 PgC/century (i.e., a long-term average ≤ 0.1 PgC yr-1). This is broadly equivalent to an atmospheric CO2 drift of less than about 5 ppm/century. We suggest that a drift smaller than this value is highly desirable but this value is a guideline. Exceeding this drift in the control run may preclude a model from being included in a C4MIP analysis, but we would expect that decision to be made on a case-by-case basis. For example, a large ocean drift in a concentration-driven experiment may not preclude analysis of land carbon fluxes and vice-versa. We also stress that being within these drifts is a minimum but not necessarily sufficient quality condition. Regional patterns and drifts of stores and fluxes will also be assessed and depending on the analysis may preclude inclusion of a given model's results.

For simulations of carbon isotopes, spin-up times of many thousands of years or the use of an equivalent fast spin-up technique may be required to eliminate drift, particularly for carbon-14 in ocean carbon and soil carbon. The spin-up technique is left to the modellers' discretion.

The pre-industrial control run (piControl) is a required simulation of the CMIP DECK, and a prerequisite simulation for participating in C4MIP. The run begins from a spun-up state as described above and all forcings should continue to be applied as per the spin-up. The global land and ocean carbon stores should not drift by more than 10 PgC/century each. The length of the pre-industrial control run should be at least equal to any simulation for which it will serve as the control simulation thereby allowing correction for model drift. The piControl run must be run for both concentration-driven and emission-driven configurations of the model. In both cases all forcings should be held constant at pre-industrial levels as described in the CMIP DECK documentation. The only difference between concentration-driven and emission-driven control runs is that the emission-driven simulation simulates atmospheric CO2 internally in response to natural fluxes of carbon from land and ocean, whereas in the concentration-driven case atmospheric CO2 concentration is specified. No anthropogenic fossil fuel emissions of CO2 should be applied to the model during this control run, and fixed pre-industrial land-use should be imposed. The simulated atmospheric CO2 in esm-piControl should therefore remain stable, with drifts below 5 ppm/century.

The CMIP6 historical run, a CMIP6 required simulation, must be performed in both concentration-driven and emission-driven configurations for participation in C4MIP. It is expected that the historical simulation would begin from the same starting point as the pre-industrial control run (Fig. 3). This nominally is set as 1 January 1850. We note though that this neglects the small but non-zero effect of pre-1850 land-use changes (see e.g., Pongratz et al., 2009; Sentman et al., 2011). Some modelling groups might therefore opt for an earlier starting date or perform additional offline land-surface simulations in order to account for pre-1850 land cover change. This would mean though that the control and historical simulations begin from different states and with different trends and this should therefore be very clearly documented. The protocol for the historical simulation is documented in detail in the CMIP6 paper (Eyring et al., 2016a). Here we stress the need for the emission-driven historical run (esm-hist) to also be performed as an “entry card” for C4MIP. The only difference between concentration-driven and emission-driven simulations is the treatment of atmospheric CO2. All other forcings must be identical in both simulations. The concentration-driven simulation will use historical atmospheric CO2 concentration provided by CMIP6.

The emission-driven simulation will use anthropogenic CO2 emissions documented here. Model groups have a choice over the treatment of land-use forcing as described below.

Fossil fuel emissions: CMIP6 will provide gridded, annual CO2 emissions from burning of fossil fuels, from the beginning of 1850 to the end of 2014 for the historical simulation and through to the end of 2100 for ssp5-8.5. See Sect. 3.3.1.

A concentration-driven simulation with a 1 % per year increase in atmospheric CO2 concentration beginning from pre-industrial is a required simulation of the DECK. In C4MIP there are further variants of this 1 % simulation designed to quantify the concentration-carbon and climate–carbon feedback parameters (Friedlingstein et al., 2006; Arora et al., 2013).

The tier-1 C4MIP simulation 1pctCO2-bgc requires the simulation to be repeated but with a change to the model set-up such that only the model's carbon cycle components (both land and ocean) respond to the increase in CO2, whereas the model's radiation code uses a constant, pre-industrial concentration of CO2. This simulation was previously known as “Uncoupled” in Friedlingstein et al. (2006), and was re-named “Biogeochemically coupled” by Gregory et al. (2009). All other forcings must be identical to the DECK 1pctCO2 simulation.

A tier-2 C4MIP simulation 1pctCO2-rad is the counterpart of 1pctCO2-bgc. It requires the simulation to be repeated but with a change to the model set-up such that only the model's radiation code sees the increase in CO2 and the model's carbon cycle components (both land and ocean) see a constant, pre-industrial concentration of CO2. This simulation was not performed in Friedlingstein et al. (2006), and was termed “Radiatively coupled” by Gregory et al. (2009). All other forcings must be identical to the DECK 1pctCO2 simulation. Although this simulation is in tier-2 we strongly encourage all modelling groups to perform it as the non-linearities of biogeochemical and radiative response can be large (e.g. Schwinger et al., 2014).

For models with a nitrogen cycle, there are two further 1 % simulation variants requested as C4MIP tier-2: 1pctCO2Ndep and 1pctCO2Ndep-bgc. These can be run if the model includes either land- or marine nitrogen cycle in a way that changes carbon uptake and storage. If the input of reactive nitrogen to the model will not affect the carbon cycle, then there is no need to perform these simulations. If changes in nitrogen deposition will affect either land or ocean carbon uptake then these simulations are requested. 1pctCO2Ndep and 1pctCO2Ndep-bgc are parallel to the 1pctCO2 and 1pctCO2-bgc simulations but with the addition of a time-varying deposition of reactive nitrogen (see Sect. 3.3.3).

Concentration-driven scenario simulations, which follow on from the end of the concentration-driven historical simulation, are performed under ScenarioMIP. In C4MIP we request simulations that complement some of these.

Under C4MIP tier-1, we request an emission-driven esm-ssp585 simulation, which parallels the ScenarioMIP concentration-driven SSP-5-8.5 simulation. This simulation should begin from the end point of the emissions-driven historical simulation (1 January 2015). As with the historical simulation the only difference from the concentration-driven counterpart should be the treatment of atmospheric CO2, which is simulated within the model driven by prescribed emissions. SSP8.5 gridded fossil fuel emissions will be provided as will SSP8.5 land-use forcing and land-use CO2 emissions. Models should implement these in the scenario run in exactly the same manner as they did in the emission-driven historical simulation.

Under C4MIP tier-2, we also request a biogeochemically coupled (BGC) version of the concentration-driven SSP5-8.5, ssp585-bgc and ssp585-bgcExt. As with the 1pctCO2-bgc simulation, this run should be performed with only the carbon cycle components (land and ocean) seeing the prescribed increase in atmospheric CO2. The model's radiation scheme should see fixed pre-industrial CO2. All other non-CO2 forcings should be applied in an identical way to the ScenarioMIP SSP5-8.5 and SSP5-8.5ext simulations. If possible this simulation should be extended to 2300, as should its counterpart from ScenarioMIP, as one of the priority focus areas for analysis is on long-term processes such as ocean carbon and heat uptake and permafrost loss (e.g., Randerson et al., 2015).

For concentration-driven simulations, atmospheric CO2 should be prescribed as a globally well-mixed value provided by CMIP6. The CMIP6 paper (Eyring et al., 2016a) and a range of papers in the GMD CMIP6 Special Issue will document the forcings in more detail. The data will be made available from the CMIP6 and PCMDI webpages (http://www.wcrp-climate.org/wgcm-cmip/wgcm-cmip6, https://pcmdi.llnl.gov/search/input4mips). For emissions-driven simulations, atmospheric CO2 should be simulated prognostically by the model. External boundary conditions of anthropogenic CO2 emissions will be provided and should be used as follows:

In esmPIcontrol, the emissions-driven control run, atmospheric CO2 should be simulated by the model but no external emissions should be added during this simulation.

LULCC affects climate via two aspects in CMIP6 simulations. In both concentration-driven and emission-driven simulations LULCC alters the distribution of vegetation covering the land surface, with consequences for the exchange of heat, water, and momentum with the atmosphere. Its effects on terrestrial carbon stocks allow us to infer LULCC emissions, more accurately labelled the “et LULCC flux” (Brovkin et al., 2013). In emission-driven simulations the net LULCC flux influences the atmospheric CO2 concentration, contributing to subsequent carbon cycle feedbacks (e.g., Strassmann et al., 2008; Arora and Boer, 2010; Pongratz et al., 2014).

The LULCC forcing for the historical simulations will be based on the protocol and forcing data provided by CMIP6 for the DECK and the historical CMIP6 simulations. LULCC is kept fixed at its pre-industrial state for all 1pctCO2 simulations (fully coupled, biogeochemically and radiatively coupled versions). It is essential that the biogeochemically coupled simulations required for C4MIP of the historical and future SSP simulations and their extensions to 2300 use identically the same LULCC forcing as for the parallel ScenarioMIP simulations.

Models including a nitrogen cycle are encouraged to use a consistent set of forcings of anthropogenic nitrogen deposition as drivers for the respective ocean and land biogeochemical components. Rates of speciated nitrogen deposition at the land and ocean surface are not available from observations and so need to be determined by models. C4MIP will coordinate with CCMI to provide gridded, time-varying fields of nitrogen deposition from chemistry transport models (CTMs) for use as driving inputs in C4MIP simulations (http://www.met.reading.ac.uk/ccmi/?page_id=375). This will be provided partitioned into four categories of wet or dry and oxidized or reduced N deposition velocities at the bottom of the atmosphere. If a model requires more or fewer categories or species of nitrogen deposition then it is up to the model group to produce these. When aggregating or disaggregating components of deposition the total amount of reactive nitrogen should be conserved. Inputs into the land biosphere depend on vegetation characteristics, and these aspects should be dealt with by the individual model groups.

C4MIP simulations should use N deposition fields as follows:

Pre-industrial control (piControl and esm-piControl) should use time-invariant, but spatially explicit, N deposition appropriate to 1850. This is so that there are no discontinuities in carbon pools or fluxes at the beginning of the historical simulation.

If the ESM simulates atmospheric chemistry and composition and therefore provides N deposition internally, then this can be used in place of a prescribed field of N deposition for the control, historical, and scenario simulations. However, irrespective of whether an ESM generates N deposition or not, for the 1 % idealized simulations, it is preferable to use the provided fields as anomalies, which should be added to the ESM's pre-industrial N-deposition fields.

The provided N-deposition data will cover both land and ocean, but we acknowledge that some models have their own established sources of reactive nitrogen to the oceans and to change this would require costly repeat-spinup simulations. So it is left to the model groups' discretion how to apply N deposition to the ocean. If a source other than that provided by C4MIP is used this should be documented and made available to aid analyses.

Models including carbon isotopes (δ13C and Δ14C) in land or ocean realms are encouraged to simulate and report variables relating to carbon isotopes for control, historical, and future scenario simulations.

For historical concentration-driven runs (piControl, historical and hist-bgc), atmospheric δ13CO2 and Δ14CO2 forcing based on observations will be provided (Fig. 4). The atmospheric forcing data sets will be available at the C4MIP website. We also plan to make available atmospheric forcing data for carbon isotopes for the ssp585 scenario and for other scenarios and extensions using a simple carbon cycle model, following Graven (2015).

Carbon isotopes are only requested to be simulated in land and ocean model components using the provided historical or future atmospheric forcing data sets for δ13CO2 and Δ14CO2. It is not requested that atmospheric δ13CO2 and Δ14CO2 be simulated by ESMs, even for emission-driven simulations of atmospheric CO2.

If the model requires any other external forcing not documented here, for example deposition of phosphorous, then it is at the model groups' discretion how to provide it. In the case of a model with an interactive phosphorous cycle, we recommend the forcing data are prepared in a way analogous to the nitrogen deposition described above. We recommend modelling groups to contact C4MIP for more details if this is applicable. Any additional forcings must be documented through the CMIP meta-data process or in the appropriate model description paper.

The primary aim of C4MIP is to compare the aspects of the global carbon cycle and its response to environmental changes across the participating ESMs. To achieve this objective, it is essential that all carbon stocks and fluxes are reported so that total amount of carbon in the system can be tracked and their conservation checked. To achieve this, compulsory tier-1 diagnostics have been defined that close the carbon cycle as simply as possible. Desirable tier-2 diagnostics should also be reported where possible, which allow for more detailed analysis by breaking down tier-1 output into sub-components.

Figure 5 shows the requested carbon cycle stores over land. Tier-1 variables are intended to be simple but still capture the total land carbon store. Tier-2 variables provide the same information as the tier-1 variables but in more detail. As shown in Fig. 5 the total carbon can be calculated from tier-1 variables and is not the combined sum of tier-1 and tier-2 variables.

The carbon stored in the vegetation–litter–soil system is simply represented by tier-1 variables, cVeg, cLitter, and cSoil respectively. For models that do not represent a vertical discretization of soil carbon, all soil carbon should be reported simply as cSoil. Additionally in tier-1 for models with vertically discretized soil carbon, we request output on the vertical distribution above and below 1 m depth (cSoilAbove1m, cSoilBelow1m). These should be reported in addition to cSoil, such that cSoil=cSoilAbove1m+cSoilBelow1m. The rationale for requesting this is the availability of several observation-based data sets that report soil organic matter content to 1 m depth. It is important that any evaluation of cSoil outputs against observed data sets makes use of the appropriate depth of soil in both the observations and model outputs.

A fourth pool, cProduct, represents the carbon stored in product pools (harvested wood, paper products, furniture, etc.) as a result of anthropogenic land-use change. The total carbon stored per unit area on land is then simply: cLand=cVeg+cLitter+cSoil+cProduct Some models may not explicitly simulate a litter pool distinct from their soil carbon pool. In this case cLitter should be reported as zero. We would normally expect cProduct to be non-zero in simulations that include anthropogenic land-use or land-use change. Hence, for the idealized 1 % per year increasing CO2 simulations (biogeochemically, radiatively or fully coupled) we would expect models to report cProduct = 0. For models whose land-use fluxes contribute straight to the atmosphere and/or to their litter or soil carbon pools, but not to the product pools, cProduct = 0 should also be reported for historical and scenario simulations. Obviously, for models that do not simulate the effect of LULCC on the carbon cycle, cProduct will also be expected to be zero.

Tier-2 output variables allow for more detailed breakdown and analysis of their parent carbon stores. They are sub-components of their parent tier-1 variables, and not additional stores. For example, the vegetation carbon pool can be represented by carbon in the leaf, stem, and root as well as possibly other (e.g. fruit) components. For models that report these tier-2 variables, the total amount of carbon per unit area should be identical to the tier-1 variable, i.e. cVeg=cLeaf+cStem+cRoot+cOther. The same applies for the litter carbon pool, which is requested to be broken down into coarse woody debris (cLitterCWD) and above- and below-surface litter (cLitterSurf, cLitterSubSurf) pools. When a model has a continuous profile of litter with depth, take above and below 10 cm as the definition of above and below the surface. CWD here is assumed to be on the surface.

For CMIP5 the soil carbon pool was requested to be divided into components with fast, medium, and slow turnover timescales. However, this distinction was not found useful by the community and as a result was not used in many analyses. For CMIP6, we request a breakdown in two different ways (Fig. 5b). First, models with a vertical structure to their soil carbon are requested to report total soil carbon for each soil layer. In the same way as soil moisture or temperature, this should be reported as a multi-level output, cSoilLevels. As the structure for this may vary between models, it is essential that the model is thoroughly documented. The sum of soil carbon over all cSoilLevels should be identical to the total cSoil tier-1 variable.

Most soil carbon models represent multiple soil carbon pools (such as fast or slow turnover, or decomposable and resistant organic material). In order to be able to diagnose and evaluate the turnover rates of carbon within the terrestrial system, we make a second tier 2 request to report individual soil carbon pools (Fig. 5b, lower panel). It is also required to report the turnover rate (tSoilPools: defined as 1/residence time) for each pool. The pool-flux structure of each model should be fully documented in its model description paper. This output will enable reduced complexity approaches (e.g. Xia et al., 2013) to recreate and analyse the soil carbon dynamics within each model. The sum of soil carbon over all cSoilPools should be identical to the total cSoil tier-1 variable.

A final tier-2 breakdown is required to report the main stores and fluxes separately for different land cover types. The LUMIP data request (Lawrence et al., 2016) requests carbon pools and fluxes for four land cover types: crop, pasture, primary and secondary land (combined as one tile), and urban. For C4MIP we additionally request a breakdown of carbon pools and fluxes within “primary and secondary” land onto tree, shrub, and grass separately. Section 4.1.4 describes the C4MIP requested output of land cover fractions. Carbon pools (cVeg, cLitter and cSoil) and fluxes (gpp, npp, ra, rh) are therefore requested on the treeFrac, shrubFrac, grassFrac, cropFrac, and pastureFrac fractions shown in Fig. 11. Table 3 lists all of these requests.

Equally important to the land carbon pools are the fluxes going into and out of them, which will allow us to gain insight into how the pools have changed and why. For ease of understanding, we have adopted a convention for newly defined variables that a carbon pool is prefixed by a “c” (as in cVeg or cSoil) and a flux by an “f” (as in fLandToOcean). Some existing variables (e.g. gpp and npp) do not conform to this but are considered to be well known and do not need to be changed.

Figure 6 shows the variables requested for terrestrial carbon fluxes. Similar to land carbon pools, the objective of tier-1 fluxes is to capture the primary system behaviour, and tier-2 fluxes provide breakdown within the tier-1 fluxes allowing for a more detailed analysis. The directions of the arrows indicate the sign convention of the flux, which is considered positive in the direction in which the arrows are pointing. For example, gross primary productivity (gpp) is positive downwards indicating flux of carbon from the atmosphere to the vegetation, whereas autotrophic respiration (ra) is positive upwards indicating flux of carbon from the vegetation to the atmosphere.

Gross primary productivity is the flux of carbon from the atmosphere to the vegetation that is associated with photosynthesis. Net primary productivity (npp) represents the carbon uptake by vegetation after the autotrophic respiration (ra) costs have been taken into account (npp = gpp - ra). Both ra and npp are sub-divided into tier-2 outputs representing flux from the leaf, stem, and root, components respectively, plus also a category “other”, which should include all the components (if any) reported under cOther tier-2 carbon pool. Also, similar to land surface pools, the sum of the tier-2 fluxes must be identical to their parent tier-1 flux. npp=nppLeaf+nppStem+nppRoot+nppOther Heterotrophic respiratory flux (rh) and CO2 emissions associated with natural wildfires (fFireNat) represent carbon loss from the land carbon stores to the atmosphere. rh is requested to be sub-divided into its tier-2 components from the litter and soil pools. Similarly, fFireNat is sub-divided into fire CO2 emissions from vegetation and litter carbon pools. Note, that fFireNat should not include CO2 emissions from fires associated with anthropogenic land-use change.

Anthropogenic land-use change or land management can result in transfer of carbon out of the vegetation, litter and soil carbon pools either directly to the atmosphere (fAnthDisturb) or to the product pool. fAnthDisturb is proposed to be split into fluxes due to land-cover change (fDeforestToAtmos) or management (fHarvestToAtmos), if this distinction is made in the model. Anthropogenic fires, associated with LUC, should be included in fAnthDisturb. Fluxes into the product pool should similarly be reported as either fDeforestToProduct or fHarvestToProduct. Decomposition of carbon in the product pool represents a carbon flux back to the atmosphere (fProductDecomp).

Due to the complexity of the processes involved, especially in the treatment of land use and management, and the growing complexity in the manner in which LUC is represented in the models, it is possible that this simple framework may not be completely compatible with all models. It is simply not possible to define in advance of CMIP6 a framework that may cover every possible flux in every model. Our request is, therefore, that all fluxes of carbon are reported somewhere, in the best possible way that they may fit within the framework shown in Fig. 6, and not missed. This will ensure conservation of carbon within the reported variables.

An example of differences in model structure and processes is the manner in which litter from the vegetation pool is transferred to the soil carbon pool. Some models simulate litter fall from vegetation into the litter pool and then subsequent assimilation into the soil carbon pool. Some models may also simulate this flux directly from vegetation to soil carbon. In either case tier-2 breakdown of the litterfall flux due to senescence (normal turnover) and mortality is requested; this breakdown is expected to help to diagnose changes in turnover time of the litter and soil carbon pools.

Figure 6 also forms the basis of carbon conservation properties that must be obeyed by the reported outputs. These include the manner in which fluxes should add up and that the rate of change of carbon in carbon pools must be equal to the sum of fluxes going in and out of the pools, or equivalently changes in pools must be equal to the sum of time integral of the fluxes into and out of the pools. gpp=npp+ra dcVegdt=npp-fVegLitter-fVegSoil-fAnthDisturb-fDeforestToProduct-fHarvestToProduct-fVegFiredcLitterdt=fVegLitter-fLitterSoil-fLitterFire-rhLitterdcSoildt=fLitterSoil+fVegSoil-rhSoilfLandToOceandcProductdt=fDeforestToProduct+fHarvestToProductfProductDecomp

We define a new variable, netAtmosLandCO2Flux, which is the total flux of CO2 from the land to the atmosphere. It should encompass every flux from land to atmosphere so that the total from each model can be compared without having to know model details of which component fluxes to sum. Due to differences in naming convention, we have chosen not to call this NBP (net biome productivity). This is an essential tier-1 variable requested from all C4MIP simulations.

Figures 7 and 8 summarize the requested terrestrial nitrogen pools and flux variables from models that include a representation of terrestrial nitrogen cycle and its coupling to the terrestrial carbon cycle. The nitrogen pools are designed to parallel their corresponding carbon stores as closely as possible, giving primarily the storage of nitrogen in the vegetation (nVeg), litter (nLitter), and soil organic matter (nSoil) pools. Additionally, we are requesting mineral nitrogen in soil (nMineral), which is sub-divided into tier-2 variables representing ammonium (nMineralNH4) and nitrate (nMineralNO3) mineral nitrogen. We do not envisage much interest in the nProduct variable (nitrogen stored in anthropogenic product pools), but it is required as a tier-1 output in order to close the nitrogen budget and ensure mass conservation of analyses. There will also be likely little interest in separating nLitter into its tier-2 components nLitterCwd, nLitterSurf and nLitterSubSurf but these variables are being requested for consistency with their carbon counterparts.

Requested fluxes associated with the flow of nitrogen over land are summarized in Fig. 8 and differ more from their carbon counterparts than do the carbon and nitrogen pools. As with the pools, all fluxes should be reported somewhere in order to be able to close nitrogen cycle budget over land and ensure mass conservation of analyses. As with carbon fluxes, the sign convention of the flux is considered positive in the direction in which the arrows are pointing

Nitrogen enters the terrestrial ecosystems either through anthropogenic inputs (which can be either atmospheric deposition, fNdep, or fertilizer input fNfert) or through biological fixation (fBNF). Flows between vegetation, litter, and soil organic N pools mirror the carbon fluxes, but with additional terms that represent inorganic mineral nitrogen uptake by vegetation (fNup) and the net mineralization flux, i.e. the difference between gross mineralization and immobilization, from the dead litter and soil organic matter pools to the mineral nitrogen pool (fNnetmin). fNnetmin should be reported as positive into the nMineral pool. Negative values of fNnetmin then imply net immobilization.

The tier-1 variables that represent the loss of nitrogen from the primary terrestrial pools of vegetation, litter and soil organic matter include fluxes due to anthropogenic disturbance: either into the LUC product pool (fNproduct) or loss direct to the atmosphere fNAnthDisturb and loss from the mineral nitrogen pool (fNloss). In order to conserve nitrogen, all losses of N must be reported into one of these variables. fNloss may be further sub-divided (if represented in the model) into tier-2 outputs of gaseous loss to the atmosphere (fNgas) and loss of dissolved organic and inorganic nitrogen through leaching (fNleach), i.e. fNloss = fNgas + fNleach. If represented in the model, fNgas can be split into that due to fire and non-fire. A further breakdown of tier-2 fluxes is also requested, if available, but these do not necessarily have to add up to the tier-1 flux value: fNOx and fN2O are components (but do not necessarily have to add up to fNgas) and may be of interest for evaluation activities or coupling to atmospheric chemistry models. fNLandToOcean may be a subset of fNleach and is of interest for studying the impact of terrestrial nitrogen cycle on coastal ocean ecosystems.

While most variables representing the land surface physical state and water fluxes will likely be requested by the land surface, snow, and soil moisture model intercomparison project (LS3MIP, van den Hurk et al., 2016) and land-use model intercomparison project (LUMIP, Lawrence et al., 2016), C4MIP requests some basic land surface physical variables as well. These include soil moisture and temperature, vegetation leaf area index (LAI) and height, and basic water fluxes.

Figure 9 shows the state variables requested that characterize the physical vegetation structure (through leaf area index and vegetation height) and the physical state of the soil (through the soil moisture and temperature of a model's soil layers).

The only tier-1 state variable requested for vegetation structure is LAI, which represents the area of leaves per unit area of ground. Vegetation height may also be considered an important evaluation metric but this is requested as tier-2 variable. It is likely more useful to distinguish vegetation height by vegetation type, i.e. by tree, shrub, grass, and crop. If this distinction is not made or unavailable in a model then only the grid-averaged vegetation height may be reported.

Soil moisture and temperature are requested as tier-1 variables to be able to analyse carbon and moisture fluxes together and to identify the role of the physical state of the soil conditions on carbon stores and fluxes. The total, liquid, and frozen soil moisture contents are aggregated and disaggregated in various ways as shown in Fig. 9 and described below:

soil temperature (tsl) is requested for each model level

Figure 10 summarizes the small number of land surface hydrological fluxes being requested. As with the carbon and nitrogen fluxes the sign convention is shown by the direction of the arrows.

prveg represents precipitation intercepted by the canopy, and evspsblveg represents evaporation from the canopy leaves (including sublimation).

Figure 11 summarizes the land cover variables requested from all models. As with other requested variables, these are categorized as simpler tier-1 variables, which represent the primary land cover types, while the tier-2 variables further break down the tier-1 variables into more detail. Tier-1 land cover variables are required from all models so that the land cover is completely described. Where possible modelling groups are requested to provide the additional details through tier-2 variables. It is important that the combined totals of tier-2 variables agree with their tier-1 counterparts.

A grid cell is described in terms of vegetation fractional coverage (vegFrac), fractional coverage of bare soil (baresoilFrac), and a residual term (residualFrac) that may include fractional coverage of urban areas, sub-grid-scale lakes, and stony outcrops. For grid cells at the continental edges, a fraction of the grid cell may also be covered by open ocean/sea. The vegFrac is further subdivided into fractional coverage by trees (treeFrac), shrubs (shrubFrac), grasses (grassFrac), crops (cropFrac), and pasture (pastureFrac). Crop and pasture fractions are the same as those requested by LUMIP (Lawrence et al., 2016). Tree, shrub, and grass fractions represent additional detail within the LUMIP tile called “primary and secondary land”. All land cover must be reported, such that VegFrac+baresoilFrac+residualFrac+SeaFrac=1,treeFrac+shrubFrac+grassFrac+cropFrac+pastureFrac=VegFrac. The tier-2 land cover variables follow the separation of trees based on their leaf structure (broadleaf and needleleaf) and leaf phenology (evergreen and deciduous) as treeFracNdlEvg, treeFracNdlDcd, treeFracBdlEvg, treeFracBdlDcd. The fractional coverage of grasses, crops, and pasture is separated into C3 and C4 variants based on their photosynthetic pathway. Tier-2 totals should sum to be identical to their tier-1 counterparts. For example treeFracNdlEvg+treeFracNdlDcd+treeFracBdlEvg+treeFracBdlDcd=treeFrac,grassFracC3+grassFracC4=grassFrac.

Figure 12 shows auxiliary land cover diagnostics and fluxes that may be reported. The additional land cover types are fractions of a grid cell related to a biogeochemical process that models may specifically simulate. These include burned area (burntFractionAll) and wetland fraction (wetlandFrac). burntFractionAll is expected to include burned area from all natural and anthropogenic processes (anthropogenic fires, and land-use change and management-related fires). wetlandFrac is expected to include natural wetlands (dynamically calculated in the model or specified) including any area of rice paddies if it is explicitly represented. Both the burnt and wetland fractions must be reported as the fraction of the grid cell and not as fraction of the land or vegetation area. Where models also estimate natural methane wetland emissions from the wetland fraction these can also be reported (wetlandCH4prod) and must include emissions from rice paddies (if represented) to make methane emissions consistent with the reported wetland fraction. If models simulate methane uptake by soils then this may be reported as wetlandCH4cons. The net land-to-atmosphere methane flux is to be reported as wetlandCH4. Models that simulate methane emissions from wetlands and/or rice paddies may explicitly simulate the depth to the water table and this may also be reported as waterDpth. Positive values of waterDpth indicate that the water table is below the ground surface and negative values indicate that the water table is above the ground surface.