Uncertainties in global crop model frameworks : effects of cultivar distribution , crop management and soil handling on crop yield estimates

10 International Institute for Applied Systems Analysis, Ecosystem Services and Management Program, 2361 Laxenburg, Austria University of Chicago and ANL Computation Institute, Chicago, IL 60637, USA Columbia University Center for Climate Systems Research and NASA Goddard Institute for Space Studies, New York, NY 10025, USA 15 Potsdam Institute for Climate Impact Research, 14473 Potsdam, Germany Comenius University in Bratislava, Department of Soil Science, 842 15 Bratislava, Slovak Republic University of Maryland, Department of Geographical Sciences, College Park, MD 20742, USA Texas A&M University, Texas AgriLife Research and Extension, Temple, TX 76502, USA Eawag, Swiss Federal Institute of Aquatic Science and Technology, CH-8600 Duebendorf, Switzerland 20 University of Natural Resources and Life Sciences, Institute for Sustainable Economic Development, 1180 Vienna, Austria Soil Science and Conservation Research Institute, National Agricultural and Food Centre, 82713 Bratis lava, Slovak Republic Department of Environmental Sciences, University of Basel, Petersplatz 1, CH-4003 Basel, Switzerland 25 Karlsruhe Institute of Technology, IMK-IFU, 82467 Garmisch-Partenkirchen, Germany Laboratoire des Sciences du Climat et de l’Environnement. CEA CNRS UVSQ Orme des Merisiers, F91191 Gif-sur-Yvette, France National Center for Atmospheric Research, Earth System Laboratory, Boulder, CO 80307, USA Department of Physical Geography and Ecosystem Science, Lund University, 223 62 Lund, Sweden 30 School of Geography, Earth & Environmental Science and Birmingham Institute of Forest Research, University of Birmingham, Edgbaston, Birmingham, B15 2TT, United Kingdom National Aeronautics and Space Administration Goddard Institute for Space Studies, New York, NY 10025, USA Peking University, Sino-French Institute of Earth System Sciences, 100871 Beijing, China 35


Introduction
Global gridded crop models (GGCMs) have become major tools in recent years for agricultural climate change impact assessments (e.g.Liu et al., 2013;Balkovič et al., 2014;Elliott et al., 2014;Folberth et al., 2014;65 Rosenzweig et al., 2014;Müller et al., 2015;Deryng et al., 2016) and water consumption studies (e.g.Liu et al., 2007;Fader et al., 2010) among others.For some models, the global gridded version is a combination of (a) a field-scale crop model, or collection of algorithms used to estimate yields and externalities of crop production for a number of pixels covering a given region and (b) a model framework (MFW) that processes input data and runs the model over large regions or the globe.The often high-complexity of these field-scale models is 70 contrasted by the scarcity of spatially detailed input data available for global-scale applications of GGCMs.This requires assumptions on crop management and the use of agricultural inputs by modelers, which can differ substantially among research groups.Fertilizer application rates may, for example, be (a) based on statistics (Liu, 2009;Deryng et al., 2011;Folberth et al., 2012;Balkovič et al., 2014), (b) defined by contrasting intensification systems (Skalský et al., 2008), or (c) integrated into management coefficients, e.g. if a model 75 does not contain explicit nutrient cycling routines (Fader et al., 2010).Other management related data such as the handling of crop residues, fallow durations and tillage practices have yet to be compiled at the global scale.
Furthermore, the setup and parameterization of a GGCM depends on the evolving objectives used in its development.The Environmental Policy Integrated Climate (EPIC) field-scale model, which is the focus of this study, has been developed for assessing impacts of agricultural management on crop growth and externalities 80 like erosion rates and soil nutrient cycling (Williams et al., 1989).When it comes to long-term impacts of climate change on agricultural production, however, soil erosion and nutrient depletion can affect yield estimates over time due to nutrient deficits (Kuhn et al., 2010).To limit such effects and facilitate investigations of climate impacts alone on plant growth, the model allows for annually resetting the soil profile.This in turn eliminates interactions among crop management, soil and climate beyond the annual growing season, which can 85 be crucial for identifying effective adaptation options (Folberth et al., 2014;Folberth et al., 2016).In GGCM studies, researchers may therefore opt for one setup over another depending on the object of investigation.
Besides different options of how to handle soils, the EPIC model provides six different methods for estimating water erosion, five algorithms for estimating potential evapotranspiration (PET) as well as several subroutines and suggested parameter ranges for nutrient cycling and gas diffusion within the soil (Gerik et al., 2014).90 Hence, GGCMs, even of the same origin or model family, can exhibit substantial differences in simulated outputs due to uncertainties originating from (a) field-scale models or modelled processes, (b) input data, and (c) setup and parameterization.In order to systematically address uncertainties in GGCMs, the Inter-Sectoral Impact Model Intercomparison Project (ISI-MIP; http://www.isi-mip.org;Warszawski et al., 2014) and the Agricultural Model Intercomparison and Improvement (AgMIP) project (https://www.agmip.org;Rosenzweig et 95 al., 2013) compiled outputs from various GGCMs that had been forced with identical climate projections (Hempel et al., 2013) in order to produce a joint agricultural climate change impact assessment (Rosenzweig et al., 2014).These studies revealed that large differences remain among GGCMs even if they are based on identical field-scale models.This emphasized the importance of distinguishing between biophysical plant growth models that are used as a core of GGCMs on one hand and MFWs on the other.100 For phase 1 of AgMIP's Global Gridded Crop Modelling Intercomparison initiative (GGCMI; http://www.agmip.org/ag-grid/ggcmi),several GGCMs have been forced with identical annual fertilizer application rates, growing seasons and climate data to eliminate uncertainties resulting from these priority data (Elliott et al., 2015).Among the 14 GGCMs participating in the study, five were based on the EPIC model, namely EPIC-BOKU, EPIC-IIASA, EPIC-TAMU, GEPIC and PEPIC.The purpose of the present study is to 105 identify differences among these EPIC-based MFWs caused by input data, crop management and parameterization.This allows for identifying priorities for further harmonization in the whole GGCM ensemble in order to refine analyses of differences in biophysical algorithms globally.All MFWs were run for six crop management intensities to evaluate the impact of nutrient and water supply on differences in crop yield estimates.Here, we focus on maize in detail as a representative crop widely used in GGCM studies. 110 Complementary results are provided for wheat in Supplementary Information S4 as key findings were in close agreement for both crops.In addition, various setup options were permutated for maize between two of the GGCMs, EPIC-IIASA and GEPIC, to assess the contribution of certain setup aspects to deviations in yield estimates.Outputs from an ensemble of non-EPIC-based GGCMs participating in phase 1 of the GGCMI project that establishes both a general benchmark for global crop models, as well as the model-driven 115 uncertainty in GGCM simulations (Müller et al., 2016), are used as a reference to which to compare the EPICbased ensemble.

The Environmental Policy Integrated Climate field scale model
The EPIC model was first developed in the 1980s to assess the impacts of soil management on crop yields 120 (Williams et al., 1989).It has been updated frequently to cover e.g.effects of elevated atmospheric CO2 concentration on plant growth (Stoeckle et al., 1992), detailed soil organic matter cycling (Izaurralde et al., 2006, Izaurralde et al., 2012), and an extended number of crop types and cultivars (e.g.Kiniry et al., 1995;Gaiser et al., 2010) among others (Gassman 2004).
EPIC estimates potential biomass increase on a daily time-step based on light interception and conversion of 125 CO2 to biomass.Plant growth and phenology are calculated based on the daily accumulation of heat units.Plant growth is constrained by water and nutrient (nitrogen (N) and phosphorus (P)) deficits, adverse temperature, and aeration stress.The potential biomass gain is subsequently adjusted by the major plant growth-regulating factor on a given day to obtain the actual biomass increment.Root growth can be limited by soil strength, adverse soil temperature, and aluminum toxicity.At maturity, the model calculates crop yield based on above ground 130 biomass and an actual harvest index HIa, which is estimated within a range given by potential HI (HImax) and minimum HI under water stress (HImin).Besides plant growth and yield formation, EPIC estimates a wide range of environmental externalities, for example wind and water erosion rates, turnover and partitioning of organic carbon (OC), N and P, evapotranspiration (ET), fluxes of selected gases, and soil hydrologic processes.

EPIC model frameworks and complementary global gridded crop models included in the study 135
Different versions of the EPIC model have been implemented in the five EPIC-MFWs evaluated in the study: EPIC-BOKU, EPIC-IIASA, GEPIC, and PEPIC use EPIC v.0810, while EPIC-TAMU uses the experimental version v.1102.Main differences between the two structurally very similar model versions are outlined in Supplementary S1.2, which also provides a comparison of maize yield estimates of the two model versions for four selected sites.Besides different model versions, the selection of subroutines adds to differences in the set of 140 algorithms used for the simulations as well.E.g., EPIC allows for selecting from five estimation methods for PET, namely Baier-Robertson, Priestly-Taylor, Penman, Penman-Monteith (PM), and Hargreaves (HG).Water erosion can be estimated by one of eight methods, and there are 11 options for estimating or inputting the water content at field capacity and wilting point.In addition, individual parameters and coefficients may be adjusted by the user to adapt the model to local conditions (Gerik et al., 2014).145 Table 1 provides an overview of major subroutines and parameters selected within the five EPIC-MFWs.
Another seven GGCMs (CLM-CROP, LPJ-GUESS, LPJmL, ORCHIDEE-crop, pAPSIM, pDSSAT, and PEGASUS) contributing to GGCMI phase 1 were included to locate the EPIC ensemble within a wider ensemble of GGCMs based on different sets of algorithms.They were selected based on the contribution of outputs for at least two of three management setups: without nutrient limitation and default or fully harmonized 150 input data (fullharm; see Section 2.4) for maize or wheat.If only default or fullharm was supplied, the corresponding other scenario was used as a substitute to keep the number of GGCMs across management scenarios constant.
The following sub-sections provide key descriptions of the setups and rationales of the EPIC-based MFWs.
GGCMs, including the seven non-EPIC-based GGCMs that were used for comparison with the EPIC-MFW ensemble, are provided on the website of ISI-MIP (http://www.isimip.org)and in Müller et al. (2016).
The spatial structure of its input data is based on a regular 5 arcmin grid, which is first aggregated to homogenous response units (HRUs) based on a classification of physical characteristics (elevation, slope, soil).
The HRUs are subsequently intersected with administrative units (national borders at the global scale) that determine specific crop management parameters, which are derived from databases or socio-economic data.The 165 field-scale model is run for each of the resulting simulation units (SimU).For comparison with GGCMs running at a 0.5°x0.5°resolution, the results from the SimUs were resampled based on the pixel-weighted 5 arcmin model outputs per 0.5°x0.5°grid.Presently, the GGCM runs two nutrient management intensities, high-input and low-input agriculture with accordingly high or low fertilizer application rates.For the default simulations, outputs from the high-input runs were submitted, corresponding to non-nutrient limited yield potential with 170 default growing season assumptions.Further management data such as growing seasons have been compiled from various sources as specified in Skalský et al. (2008).

EPIC-IIASA
EPIC-IIASA has been developed in parallel to EPIC-BOKU, partly by the same researchers and shares the same spatial data structure based on HRUs and SimUs.Parameterizations and input data have been adjusted 175 throughout research projects resulting in a substantially differing setup with the major remaining communality being the use of a static soil profile (Table 1).Growing seasons have been adopted from Sacks et al. (2010) and crop-specific spatially explicit N and P application rates from Mueller et al. (2012).Focus regions of recent studies for which model setups have been adjusted are the EU (e.g.Balkovič et al., 2013) and China (Xiong et al., 2014a;Xiong et al., 2014b) besides global applications (Balkovic et al. 2014Xiong et al. 2016).180

EPIC-TAMU
EPIC-TAMU follows the model development and implementation of EPIC, version 1102, which accounts for C and N stocks and flows in managed terrestrial ecosystems (Izaurralde et al. 2012).As in EPIC v. 0810, the coupled C and N model in EPIC-TAMU follows the conceptual pool structure of the Century model (Izaurralde et al., 2006).Mineralization and immobilization of C and N also follows the approach in Century but a recent 185 option has been added to describe C and N of microbial biomass following the approach used in the Phoenix model (McGill et al., 1981).The EPIC-TAMU version also contains algorithms to model the effects of biochar additions on crop productivity, soil pH, and cation exchange capacity (Lychuk et al., 2015).Other developments include a mechanistic model to describe microbial denitrification and the corresponding feedback on decomposition (Izaurralde et al., 2012).190 EPIC-TAMU has primarily been used for field-scale and regional-scale simulations (Gelfand et al., 2013;Zhang et al., 2015).It has been adapted with minimal changes for use as part of the AgMIP GGCMI project, and has Biogeosciences Discuss., doi:10.5194/bg-2016-527,2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.
otherwise not been previously used for global simulations.As a result, no default simulations (see Section 2.4) were produced.To keep the number of EPIC-MFWs in evaluations across management scenarios constant, the fully harmonized setup was also used as default.195

GEPIC
The GEPIC MFW was originally developed for studies of global crop-water relations (Liu et al., 2007).In its present version, it uses input data for planting dates, growing season length, P fertilizer application rates, and cultivar distributions besides the original input data elevation, slope, country/region, N fertilizer application rates, and irrigation water management (Folberth et al., 2012).Default fertilizer inputs are mainly based on the 200 FertiStat database (FAO, 2007) and have been extrapolated for maize based on the human development index (HDI) for countries lacking data.
More recently, the MFW has been setup for applications in sub-Saharan Africa based on a regional calibration (Gaiser et al., 2010) and thoroughly evaluated in various studies (e.g.Folberth et al., 2012;Folberth et al., 2014;van der Velde et al., 2014).The authors found that when using dynamic soil profiles in the setup, the model 205 reproduces yields around the year 2000 well after a spin-up of 30 years (Folberth et al., 2012).Extending the simulation period may result in erosion of the whole soil profile at some point or complete nutrient depletion in grid cells that lack fertilizer inputs.To avoid resulting detrimental effects on crop yields and unrealistically long monocultures, the model is run for each decade of the study period separately, which aims at mimicking fallow rotation with an average cultivation period of 40 years and complete recovery of the soil profile afterwards (see 210 Figure S1-1).

PEPIC
PEPIC is a global EPIC-MFW developed at the Swiss Federal Institute of Aquatic Science and Technology (Eawag) initially based on GEPIC, which had been developed at the same institute.Hence, the two MFWs have similar features in software design and default input data.However, one of the main purposes of PEPIC was to 215 develop a fully free tool, which can be used without any software license.Therefore, unlike GEPIC, PEPIC was compiled by a free computer language, Python.In addition, the parameterization and setup has been adjusted in large parts (Table 1) to match focus research purposes.It was initially developed to investigate the impacts of different PET methods on crop-water relations (Liu et al., 2016a).Presently, applications of PEPIC focus on assessing trade-offs between crop yields and nutrient losses, e.g.N, in the context of global agricultural 220 intensification (Liu et al., 2016b).For such assessments, N was applied three times during the whole season following a fixed schedule (Liu et al., 2016b).Besides the parameterization, this is a single major difference compared to the other four EPIC-MFWs, which used automatic N fertilization based on plant nutrient requirements.

Common input data
Climate forcing data based on the WFDEI GPCC dataset (Weedon et al., 2014)  solar radiation from ERA-interim (Dee et al., 2011) and precipitation from GPCC (Schneider et al., 2013).All 230 EPIC-based MFWs used soil data from the ISRIC-WISE database (Batjes, 2006) mapped to the Digital Soil Map of the World (FAO, 1995).For EPIC-BOKU and EPIC-IIASA, the originally approx.5000 soil profiles had been reduced to 120 based on a classification of key variables (Skalský et al., 2008) and soil hydraulic parameters not provided in the WISE databases were complemented according to Schaap and Bouten (1996) and Wösten and Van Genuchten (1988).235 For the harmonized runs, nutrient application rates for N and P were based on crop-specific data from Mueller et al. (2012) to which manure application rates had been added proportionally.Separate planting dates and growing season lengths for rainfed and irrigated management were based on Sacks et al. (2010), complemented by gap filling with data from the MIRCA2000 dataset (Portmann et al., 2010) and the LPJmL model (Waha et al., 2012).Both datasets were provided by the GGCMI project (Elliott et al., 2015).Default runs (see Sect. 2.4; 240 Table 2) were carried out using individual fertilizer and growing season databases within each MFW.

Crop management scenarios
Six crop management scenarios (Table 2) were simulated to quantify differences among MFWs based on three steps of growing season and nutrient supply harmonization.Each of these three scenarios was simulated with two water management scenarios, rainfed only or sufficiently irrigated.245 The default scenario represents each research group's assumptions on fertilizer application rates and growing seasons (see Sect. 2.2).It serves for evaluating differences among MFWs if only climate data are harmonized.
The fully harmonized (fullharm) setup allows for identifying remaining differences if annual nutrient application rates and growing seasons are harmonized using the input data described in Sect.2.3.The fully harmonized setup with sufficient nutrient application (harm-suffN, referred to as harmnon in the simulation protocol 250 published by Elliott et al. 2015) virtually eliminates plant nutrient deficits and consequently impacts of soil nutrient dynamics.This is expected to minimize differences among MFWs resulting from the setup of fertilizer application and soil nutrient cycling.
Irrigation water is applied in all MFWs up to a sufficient amount automatically based on plant water requirement in the irrigated management scenarios.The application takes place based on varying thresholds (1-255 20% plant water stress, see Supplementary S1) in the MFWs.

Geographic distributions of maize cultivars
Crop cultivars are here considered to be defined by HImin, HImax, and optimal temperature ranges only, whereas 260 the heat unit requirement is prescribed by growing season input data.Between one and four different maize cultivars were planted within each MFW (Figure 1; Table S1-1).EPIC-IIASA uses four cultivars in its default setup (Figure 1a) that are attributed to major world regions based on climatic and economic characteristics.The same geographic distribution of cultivars was used for EPIC-IIASA in the harmonized setup scenarios except that the early and late maturing high-yielding cultivars 1 and 3 were merged as growing season length was 265 defined according to common input data sets (see Sect. 2.4).EPIC-TAMU (Figure 1b) plants one high-yielding and one low-yielding variety.The latter was attributed to countries, in which maize yields stagnated or Biogeosciences Discuss., doi:10.5194/bg-2016-527,2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.decreased within the past decades according to Ray et al. (2012), the first to all other.The same two maize cultivars were distributed in the GEPIC and PEPIC MFWs (Figure 1c) based on the HDI.The high-yielding variety is planted in all countries with HDI≥80, which corresponds to high development.EPIC-BOKU used the 270 high-yielding variety in all grid cells (Figure 1d).

Permutation of setup options for GEPIC and EPIC-IIASA
To better identify the importance of single data and parameterization domains within the MFWs, selected 275 aspects of model setups were exchanged between EPIC-IIASA and GEPIC.These were grouped into the domains of cultivar distribution (Cult), soil data (SoilD), soil handling (SoilP), crop management (Manage) and miscellaneous coefficients (Coeff) (Table 3).The GEPIC model was run with all 32 resulting setup combinations using the land mask of EPIC-IIASA to ensure consistency.The evaluation focused on rainfed yield estimates as these cover the whole range of uncertainty impacts.Yield estimates based on sufficient 280 irrigation were used to identify differences based on soil hydrology and other factors affecting water limitations to plant growth.

Reported yield data and evaluation 285
Unless otherwise specified, all results were evaluated for the period 1980-2009, which corresponds to the span of the climate data across growing seasons in both hemispheres.Annual national average crop yields from the FAOSTAT database (FAO, 2014) were used for assessing model performance.Reported yields were de-trended using a 5-year moving mean average in order to remove trends in yields due to changes in technology and management.Model performance, which served here solely for comparing differences in model skills in relation 290 to differences in setups, was evaluated using time series correlation (tscorr) between national average simulated and reported yields using Pearson's product moment correlation coefficient r.A tscorr of r>0.5 (corresponding to more than 50% of variance explained by the model) was selected as threshold for good model performance.
The coefficient of variation (CV) was used as a metric of deviation among yield estimates from different

MFWs. 295
National average yields (YDav) were calculated from simulated rainfed and irrigated yields in each grid cell and the respective rainfed and irrigated harvested areas (Portmann et al., 2010) according to: where YDav,c is the national average yield in country c, YDi,g is yield under irrigated conditions in grid cell g, YDr,g is yield under rainfed conditions in grid cell g, HAi,g is irrigated area in grid cell g, and HAr,g is rain fed area in each grid cell g, and m is the number of grid cells in country c.We acknowledge the uncertainty introduced from spatial aggregation (Porwollik et al., 2016) but as the focus is on a comparison of MFWs, we consider this to be of minor importance here.305 All evaluations were carried out with the statistics software R (R Development Core Team, 2008) using the packages ggplot2 (Wickham, 2009) and the heatmap.2function of gplots (Warnes et al., 2016) in a modified version by Müller et al. (2016).

Global average maize yields 310
If the GGCMs are run in their default setups, global average simulated maize yields differ by up to 124% annually (mean 95%) using the lowest estimate as a reference (Figure 2a; Table S2-1).This is mainly due to very high yield estimates from EPIC-BOKU of around 8 t ha -1 , while the other EPIC-MFWs have yield estimates of around 4-6 t ha -1 .The ranges decrease to 55% if harmonized planting dates and fertilizer application rates are used (Figure 2b) and further to 26% with sufficient nutrient supply (Figure 2c).In addition, 315 the order of EPIC-MFWs mean biases changes.While EPIC-BOKU has the highest average default yield estimates, the highest yields simulated given sufficient nutrient supply and harmonized growing seasons are from EPIC-IIASA and EPIC-TAMU.
EPIC-BOKU and PEPIC exhibit declining yield trends over the simulation period, whereas the other EPIC-MFWs do not exhibit any trend.For PEPIC this is most pronounced in the default setup, for EPIC-BOKU in the 320 harm-suffN scenario, in which it initially provides the highest yield and at the end of the simulation period one of the lowest.The inter-annual yield variability shows very similar patterns among all MFWs.The whole EPIC-MFW ensemble indicates a peak in global average yield in 2004, which is also apparent in the reported data and the non-EPIC-based GGCMs (Figure S2-1).
The continuous decrease in the spread among EPIC-MFWs with increasing level of harmonization and 325 elimination of nutrient limitations is contrasted by an increasing spread for the non-EPIC-based GGCMs (Figure S2-1), comprising the models CLM-CROP, LPJ-GUESS, LPJmL, pAPSIM, pDSSAT, and PEGASUS as ORCHIDEE-crop provided data for the harm-suffN scenario for wheat only (see Supplementary Information S4).This is mainly driven by very high yield estimates by two of the non-EPIC-based GGCMs and very low estimates by one, whereas three of the non-EPIC-based GGCMs estimate yield potentials at a similar level as 330 the EPIC ensemble.Besides this increase in spread among MFWs, the range of yields is in the fullharm scenario among the non-EPIC-based GGCMs about three times as large as among the EPIC-MFWs.

Spatial patterns of differences in maize yields among model frameworks 335
No clear relationship between maize cultivar distribution and differences in maize yields could be identified (Figure S2-6).Spatially, the deviation of maize yield estimates among the MFWs is largest with the default setups in tropical and arid regions with CVs of up to 224% and CV ≥50% in >44% of all grid cells (Figure 3;  S2-2).The most distinct differences were found in sub-Saharan Africa, particularly the countries of Angola and Mozambique.Large differences with CVs around 100% occur also in South America, Harmonizing fertilizer and growing season data substantially reduces the CV among models to ≤64% under rainfed and ≤54% under irrigated conditions in 75% of all grid cells (Table S2-2).Spatial patterns remain largely similar to those found for the default managements.However, the CV increases in a few regions after 350 harmonization.E.g., in the low fertilizer input region Western Russia EPIC-IIASA, which plants here a cultivar adapted to colder climates, provides high yields in comparison to the other EPIC-MFWs, which raises CV from around 30-40% in the default setups to often 50-60% in the fullharm ensemble.Also, in the semi-arid region of North-eastern Brazil, EPIC-IIASA provides very high yields, while EPIC-BOKU and EPIC-TAMU do so in Southern China.Both regions have reported growing season lengths >200 days, which may amplify differences 355 in nutrient application, soil hydrology, and planted cultivars.
The application of sufficient nutrients further reduces differences among MFWs.Under irrigated conditions, the CV is now in 50% of the grid cells ≤25% and under rainfed conditions ≤29%.Differences among model frameworks decrease especially in regions with low or moderate reported fertilizer application rates such as India, sub-Saharan Africa, and South America.Large CVs remain in the same regions as in the fullharm setup, 360 foremost in Southern China.

Impact of fertilizer supply on maize yield deviations
The patterns observed in Figure 3 indicate that yield estimates differ especially in low-input regions.365 Accordingly, differences in yield estimates decrease among models with increasing N application rates (Figure 4).This relationship is more evident under irrigated than under rainfed conditions.The CV is on average at about 60-85% in all climate regions at very low N application levels and highest in arid regions under irrigated conditions.It decreases on average to about 21-23% in arid and temperate regions and 16-17% in tropical and cold regions at applications rates above 150 kg N ha -1 yr -1 .Rainfed cultivation virtually eliminates nutrient 370 application rates as a driver for differences among models in (semi-) arid climates and dampens the effect of nutrients substantially in temperate climates.Rainfed cultivation in cold and tropical regions in contrast shows similar patterns to that with sufficient irrigation, but with higher CV values especially at higher nutrient application rates.

Impact of single setup options on maize yield estimates
The permutation of model setup options among the MFWs EPIC-IIASA and GEPIC (Table 3) for the fullharm scenario under rainfed conditions shows general principles and interdependences of several of these parameterization domains (Figure 5; see Figure S2-4 for relative differences among scenarios).For a description 380 of the coefficient groups compared here see section 2.6 above.A leading "g" in the coefficient name in Figure 5 and the text indicates that the parameters are taken from the GEPIC MFW, whereas a leading "e" indicates that the parameters are from the EPIC-IIASA MFW.A summary of the key findings is provided in Table S2-3.If a static soil profile is assumed (eCoeff, columns one and three in Figure 5), there is hardly any impact of the soil handling (SoilP).In addition, the residue management (line color) hardly affects yield levels, in contrast to a 385 dynamic soil profile (gCoeff) where the management of GEPIC (gManage) allows for higher yields at any point.
If sufficient irrigation water is supplied (Figure S2-3), however, a difference in yield estimates is apparent also for eCoeff runs.This can be linked to the higher sensitivity of the model for nutrient limitations if sufficient water is supplied and faster turnover of plant residue at adequate soil humidity.Besides these nutrient-related effects, selecting the eCoeff setup domain results in a positive yield trend, which must be attributed to 390 precipitation in the climate input data as there are no transient effects of soil management and the trend cannot be observed in the irrigated management.Selecting gCoeff, in contrast, results in a negative yield trend, larger amplitude in inter-annual yield variability, and an apparent effect of plant residue management.This trend is more pronounced if the model runs are carried out fully transiently (eSoilP) rather than on a decadal basis (gSoilP).395 Figure 5 The choice of soil data (SoilD; rows one and two vs. three and four) has an impact on inter-annual yield dynamics, especially if the parameterization and soil handling of EPIC-IIASA (column one) is used.As the 400 effect of soil data is less apparent in permutations with a stronger influence of soil nutrient dynamics (e.g.gCoeff; columns two and three), the difference can be traced to differences in soil hydrology.This is supported by the fact that differences in yield dynamics are less pronounced if sufficient irrigation water is supplied (Figure S2-3).However, the identification of an exact driver will have to be subject to detailed evaluations at the site scale altering one parameter at a time.405 The choice of cultivar distribution foremost affects the magnitude of yields with about 20% higher yields for the distribution used in EPIC-IIASA (rows one and three vs. two and four in Figure 5).This corresponds closely to the differences in HImax between the cultivars (Table S1-1) and their distribution across the global maize harvested area (Figure 1a, c).In addition, there is an effect on inter-annual yield dynamics if sufficient irrigation water is supplied and the parameterization and soil handling of GEPIC is used column 410 four).In this case, the maize distribution of EPIC-IIASA shows larger inter-annual variability with higher peaks for example in the early 1990s and 2000s.

Differences in the performance of model frameworks and setups
The time series correlation (tscorr) between annual national average yields and detrended reported yields shows a mixed picture of MFW performance for maize (Figure 6a).The ranking order of the EPIC-MFWs by number EPIC-IIASA (10), EPIC-BOKU (7), GEPIC (5) and PEPIC (4).However, in various countries more than one MFW has r>0.5 (Figure 6b).When considering all incidences in which an MFW has r>0.5 (see also Figure S3-420 1), the gap among MFWs narrows down with EPIC-TAMU (30), EPIC-IIASA (25), EPIC-BOKU (23), GEPIC (22),and PEPIC (22).Spatially, none of the MFWs exhibits a distinctively high performance in any part of the world defined by management intensity or climate conditions.In general, the performance is poorest in tropical regions around the equator.Out of the ten countries with the largest maize production volumes (Figure S3-2), seven have a tscorr with r>0.5 for at least one MFW, and all MFWs exhibit a high performance in the USA and 425 France.
Figure 6 Testing the permutations of setup options for performance in reproducing reported maize yields does not reveal 430 a strong tendency towards one of the two MFWs nor a spatial pattern of optimal setups (Figure 6c; Figure S3-3).
Various countries show the best performance if mostly the setup options from EPIC-IIASA are used in combination with the coefficients and/or management of GEPIC (e.g.Brazil, Canada, Mexico, or Zambia).In turn, an optimal setup mostly based on GEPIC combined with management operations or soil data of EPIC-IIASA can be found e.g. in Greece, Italy or Japan.In most countries, the highest performance is achieved with a 435 rather even mix of setup options from both MFWs.Thereby, the setup components soil data, soil handling, and coefficients of EPIC-IIASA have a slightly higher frequency of 53-55% and the management a slightly lower with 49% in the optimally performing setups (Figure S3-3).The cultivar distribution of EPIC-IIASA in contrast shows the best performance in 72% of all countries.However, as the cultivar distributions of both MFWs are identical in various countries, also the GEPIC cultivar distribution shows the best performance in 60% of all 440 countries.These figures are similar if only countries with tscorr r>0.5 are considered, but the frequency of better performing EPIC-IIASA setup domains increases for coefficients to 63% and the performance of cultivar distributions changes to 75% for EPIC-IIASA and 54% for GEPIC.
Still, often various setup permutations show a comparably good performance in most countries (Figure 6d; Figure S3-3).Out of the countries with tscorr r>0.5, only Benin, Lithuania, and Peru have a high agreement for 445 only one or two setups.In particular European countries, south-east African countries, the USA and Thailand exhibit a good agreement with often more than 20 and up to all 32 setup variations.For single countries, certain setup domains or their combinations can be identified as affecting model performance (Figure S3-3; Figure S3-4).For example, in Argentina, the coefficients of GEPIC cause a lower performance compared to each equivalent simulation with EPIC-IIASA coefficients.This is the opposite, although to a lesser extent, in China.450 In India, the coefficients of GEPIC combined with the management of EPIC-IIASA lead to lower model performance especially if combined with the cultivar distribution of EPIC-IIASA.

Crop distributions and parameterizations in the model frameworks
The analyses performed here do not allow for identifying optimal maize cultivar distributions based on the 455 EPIC-MFWs yield estimates and observations.This is due to the wide range of differences among EPIC-MFW setups and will hence have to be subject of a focus study.From a general point of view, each of the four approaches for spatially distributing maize cultivars (in terms of optimum temperature ranges and HI as defined in Table S1-1) globally used by the EPIC-MFWs follows logical rules but with differing rationales.The zoning approach based on agro-climatic and economic regions used in EPIC-IIASA (Figure 1a) accounts for climatic 460 adaptation of cultivars as well as access to improved seeds.Using reported long-term yield dynamics to distribute maize cultivars as in EPIC-TAMU (Figure 1b) takes processes in agricultural development into account.The HDI used in GEPIC and PEPIC (Figure 1c) reflects socio-economic means for investing in improved inputs and access to knowledge.Finally, planting the same cultivar globally as done in EPIC-BOKU (Figure 1d) ensures high spatial comparability of plant growth processes in the model.However, none of the 465 approaches can be considered to fully reflect the actual distribution of maize cultivars, including the challenging problem of sub-grid heterogeneity.Already the fact that there are strongly differing domestic agricultural systems in large transition countries like Brazil, China and India implies the requirement for a more nuanced approach of distributing cultivars.As neither the national, nor the subnational, socio-economic situation can fully serve as a proxy for agricultural management intensities, a global characterization of agricultural 470 production systems will be required to develop a valid dataset for current maize cultivar distributions.Such a baseline distribution, along with characterization of the drivers will also be a crucial first step for developing realistic projections of future cultivar distribution under environmental change.

General setup components 475
The largest differences for global average maize yield estimates among EPIC-MFWs in their default setups (Figure 2) result from the planting of a high-yielding cultivar and uniformly high fertilizer supply in EPIC-BOKU, contrasted by present fertilizer inputs and expansive planting of a low-yielding cultivar in GEPIC and PEPIC.The inclusion of a dynamic soil profile and soil erosion in the latter two further adds to this.EPIC-IIASA and EPIC-TAMU have lower extents for planting the low-yielding cultivar and use a static soil profile 480 and/or full water erosion control in their setups.In contrast to the other EPIC-MFWs, PEPIC exhibits a negative yield trend over time.This can be attributed to the use of fixed fertilizer application schedule in PEPIC (see section 2.2.5) in contrast to automatic application based on plant requirements in the other EPIC-MFWs, which typically increase nutrient losses, and the inclusion of soil erosion combined with fully transient model runs.
Both of which have been shown to affect yield trends in the model before (Liu et al., 2016b).The fact that the 485 mean bias among EPIC-MFWs becomes very small with sufficient nutrient supply highlights that nutrient supply is the main driver for differences in mean model bias.However, the remaining differences in inter-annual The facts that the non-EPIC-based GGCM ensemble shows a relative increase in the mean range of yield estimates with increasing level of harmonization and that the spread among non-EPIC-based GGCMs in the fullharm scenario is about three times the spread among the EPIC-based MFWs (Figure S2-1) highlight that the uncertainty introduced by differences in parameterization and input data in the same model is still far lower than uncertainties related to conceptual differences among models.These have been characterized in Müller et al. 495 (2016) and on https://www.isimip.org/impactmodels.General differences such as the inclusion of certain climate variables and employment of contrasting representations of photosynthesis and biomass partitioning may have less of an effect on this phenomenon.More importantly, the models take soil characteristics (and hence their limiting effects) into account at varying levels of detail, have cultivars based on plant functional types or specific crop cultivars, and have explicit representations of nutrient supply or include these through 500 aggregated management coefficients.In addition, several of the GGCMs have been calibrated to reproduce present yield levels in their default setups, either by altering such management coefficients, or by fitting the crop phenology.In these cases, changing growing season and fertilizer input data typically results in new combinations of these fitted variables and agronomic conditions, which can cause the observed yield estimates far above or below the reported levels despite the good fit in the default setups.The EPIC model in contrast was 505 specifically developed to investigate impacts of crop management on yields and due to the same set of algorithms employed in the EPIC-MFWs an increasingly good fit could be expected with increasing level of harmonization.

Spatial differences among model frameworks and effects on yield estimates
Spatially, low yields and tropical climate are characteristic of regions in which the EPIC-MFWs show large 510 differences in maize yield estimates in their default setups (Figure 3).The large CV is in part a result of the low yield estimates in EPIC-IIASA, EPIC-TAMU, GEPIC and PEPIC, resulting in a low mean yield, but is mainly due to the estimation of yield potential by EPIC-BOKU (Figure 2a).Besides fertilizer inputs, assumptions about growing seasons differ more strongly in tropic and warm arid regions, where cropping is often possible throughout the whole year, in contrast to temperate and cold climates which usually provide only a narrow 515 growing-season window.The fact that irrigation increases differences among EPIC-MFWs in parts of sub-Saharan Africa and India is mostly due to further increases of non-nutrient-limited yield potential estimates in EPIC-BOKU.An additional minor effect is the elimination of plant water stress, which amplifies differences among the other EPIC-MFWs due to a higher sensitivity of EPIC to differences in nutrient supply (see Sect. 2.1).This is most pronounced in low-input regions, where nutrient supply depends on the parameterization of 520 soil nutrient cycling and crop residue management.
The harmonization of annual nutrient application rates and growing season dates decreases the differences among EPIC-MFWs substantially.The main driver is the adoption of reported fertilizer application rates in EPIC-BOKU.The effect is hence largest in present low-input regions such as sub-Saharan Africa, South America and parts of Asia.The larger differences under rainfed conditions -mostly in (semi-)arid regions -may 525 be attributed to differences in soil hydrology based on differing soil data, tillage and residue management, and the parameterization of soil and runoff processes, all of which affect plant water availability (see Folberth et al., 2016).Smaller differences remain in Europe, the USA, Oceania, Japan and southern South America where the predominant use of high-yielding maize cultivars is reflected in all MFWs.The application of sufficient nutrients mostly leaves CV<40% for maize yields among models.The remaining 530 difference can to a large extent be explained by the use of different maize cultivars, which differ by about 50% in genetic yield potential if the lower value is used as a base.As nutrient supply from the soil is virtually eliminated in this management scenario, factors affecting plant available water, such as estimation routines for PET, runoff, and soil hydrology (Table 1) also contribute to remaining uncertainties.If sufficient irrigation water is supplied, major differences in setups remain in fertilizer (threshold for automatic application or timing 535 of rigid application) and irrigation water (threshold for automatic application and maximum volumes) application.
Although the CV among EPIC-MFWs decreases substantially with harmonization of growing seasons and fertilizer rates, and further with application of sufficient nutrients, the global deviation among crop yield estimates is still too large to allow for a detailed quantification or ranking of drivers in these deviations.For 540 example, low yields may occur as well if a high-yielding cultivar is planted and soil nutrients are depleted (e.g. from long-term cultivation or soil erosion) or if a low-yielding cultivar is planted and moderate amounts of fertilizer are applied (see Folberth et al., 2012).Substantial differences remain among the model setups when it comes to processed soil data, soil nutrient cycling and the handling of crop residues after harvest.All of these are major determinants for nutrient availability if exogenous nutrient application is low or completely lacking.545 The remaining differences among yield estimates highlight that a further harmonization may need to prioritize the handling of fertilizer application (i.e.thresholds for automatic application or timing for fixed rates) in the MFWs besides the parameterization of soil nutrient dynamics, soil data itself and the handling of crop residues.
Recently, the substantial impact of soil type selection in the EPIC framework GEPIC has been evaluated in detail by Folberth et al. (2016) showing that soil data can drive absolute yield levels as well as inter-annual yield 550 dynamics and the crop's sensitivity to adverse climate conditions.While a harmonization of soil data is difficult across a range of GGCMs with highly differing conceptual treatment of soil processes and required variables, phase 2 of the GGCMI project prescribes a rigid fertilizer application schedule to eliminate this source of uncertainty.
The largest differences remain in the tropics, thus highlighting the need for more extensive model evaluations 555 and data collection in these regions.Continuous high temperatures and humidity levels in the tropics allow for a rapid turnover of organic material like plant residue.At the same time, mineral nutrient application rates are often low in these regions, i.e. most of Sub-Saharan Africa and parts of South America.The parameterization of these turnover processes can here be a key factor for minimizing differences among EPIC-MFWs, which is also indicated by the large differences among EPIC-MFWs in low-input grid cells (Figure 4).Furthermore, we find 560 the distribution of maize cultivars among EPIC-MFWs differs the most in these regions.This input can be considered a priority for harmonization when it comes to reproducing absolute yield levels but -due to differences in water stress sensitivity among different cultivars -also in the estimation of inter-annual yield variability.
Concerning the cold and temperate zones, a recent study based on PEPIC has shown that the choice of the PET 565 estimation method causes the largest differences in these regions (Liu et al., 2016a).It will hence need to be tested, how an ensemble of GGCMs reacts to different choices of PET estimation methods.

Impact of single setup domains on maize yield estimates
The comparison of EPIC-IIASA and GEPIC shows that both EPIC-MFWs are capable of producing fairly stable yield levels over time within their respective setups, which is essential for estimating climate change impacts on 570 crop yields without bias from trends in soil quality.However, they differ substantially in magnitudes of yields as well as in inter-annual yield variability.The latter can be traced to (a) the use of static vs dynamic soil profiles, (b) cultivars with different water stress sensitivity, and most importantly (c) the parameterization and choices of subroutines as specified in Table 1.
Both EPIC-MFWs use the Hargreaves method (Hargreaves and Samani, 1985) to estimate PET, but with 575 different parameterizations.GEPIC uses the original exponential coefficient and a linear coefficient that had been adjusted by the EPIC developers for prior studies.Vice versa, EPIC-IIASA uses the original value for the linear coefficient and an adjusted exponential coefficient.Neither of the two parameterizations can be considered more favorable than the other for global-scale applications, as PET estimation methods in general require a local calibration (e.g.Gavilán et al., 2006).This is presently not feasible at the global scale due to lack 580 of validation data.The selected parameter values cause higher PET in GEPIC at low to moderate daily temperature ranges, especially at high average temperatures and higher PET in EPIC-IIASA at high diurnal temperature ranges.An effect is that GEPIC has a higher likeliness of water stress in most climate regions, except for deserts and mountain ranges, where large differences in day and night time temperatures are common.In addition, the soil evaporation coefficient had been lowered in EPIC-IIASA from its default value, 585 which decreases the fraction of ET evaporating from soils.All these hydrologic factors contribute to differences among the EPIC-MFWs in inter-annual yield variability when comparing columns one and two or three and four of Figure 5 and Figure S2-4.
As the parameterizations of the two EPIC-MFWs were exchanged collectively, however, other parameters discussed above for the whole ensemble such as plant stress thresholds for fertilizer application and coefficients 590 for microbial turnover of soil organic matter (SOM) also play a role here.The upper limit for denitrification was set very low in GEPIC due to prior applications for sub-Saharan Africa based on a calibration of the EPIC model for the region (Gaiser et al., 2010), whereas denitrification was allowed to be high in EPIC-IIASA.The microbial decay rate, a coefficient for turnover rates of SOM, in contrast, was strongly limited in EPIC-IIASA and set to a default value corresponding to no limitations due to microbial activity in GEPIC.The latter causes 595 more rapid mineralization of nutrients from SOM.This drives differences between the two MFWs regarding residue management and nutrient cycling.Also the option of using a dynamic or a static soil profile contributes substantially to crop yield estimates.Depending on the climate and management, the static soil profile can in each year potentially provide the same amount of nutrients from SOM mineralization whereas the dynamic soil profile will experience soil nutrient depletion in low-input regions before being reset after a decade (see 600 Supplementary 1).This may cause the overall more negative yield trends in the gCoeff scenarios and contributes to differences in inter-annual yield variability due to diverging handling of soil structure and SOM across growing seasons.
The soil data used in the two models are based on the same dataset but have been aggregated to different standardized soil profiles for EPIC-IIASA as described in Sect.2.3.Still, organic carbon (OC) contents are 605 identical in >75% and within a range of ±20% in nearly 90% of all grid cells globally, which renders nutrient availability from initial SOM a minor determinant for differences in yield estimates (not shown).A major Biogeosciences Discuss., doi:10.5194/bg-2016-527, 2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.
difference is that water content at field capacity and wilting point, as well as hydraulic conductivity, are being estimated by different pedotransfer functions in the two MFWs (Table 1).This causes on average about 10% higher rainfed yield estimates in EPIC-IIASA (Figure 5, column one), an effect that is largely cancelled if 610 sufficient irrigation water is supplied (Figure S2-3 and Figure S2-5, column one), and emphasizes the importance of high resolution and quality soil data, as well as also raising questions about how to handle subgrid heterogeneity in such parameters (e.g.Folberth et al., 2016).
The cultivar distribution predominantly scales absolute yield levels.Based on HImax of the cultivars and their coverage within the two MFWs, the genetic yield potential is 19% lower in the GEPIC dataset than in EPIC-615 IIASA.This corresponds closely to the difference observed among the corresponding yield estimates (Figure 5, rows one and three vs. two and four).The HImin under water stress for cultivar 1 affects yield estimates only under severe water stress.This influences the inter-annual yield variability by causing disproportionately lower yields in extremely dry years.Apart from the HI also the temperature response of cultivar 3 used in the EPIC-IIASA setup has an impact on crop phenology and hence inter-annual yield variability.The performance when 620 using the EPIC-IIASA cultivar distribution if better in various Eurasian countries in which this cultivar is planted, such as Albania, Austria, Czech Republic, Germany, Luxembourg, Moldova, and Slovakia (Figure S3-3).However, also the default cultivar 1 used in GEPIC in the same region shows in some instances a better performance, e.g. in Belgium, Bulgaria, and Greece.

Model framework performances in estimating reported crop yields 625
As the variability in observed national yield time series is often also driven by non-weather related variability of unknown importance, GGCMs cannot be expected to explain yield variability in all countries.Müller et al. (2016) established a first benchmark for GGCM performance where the EPIC-MFWs show similar performance to the other GGCMs in that study.Here, a good agreement between estimated and reported yields was found for various major producing countries, where fertilizer inputs are usually moderate to high and stable harvest area 630 distributions within countries (as assumed in the constant global harvested areas dataset used for evaluations) can be expected In low-input regions in contrast, there is commonly little agreement with reported yields.This has various reasons such as (a) management may be very particular, e.g.slash and burn agriculture or crop rotations that are commonly not reflected in the GGCMs, (b) yields depend strongly on agro-environmental processes such as nutrient cycling, (c) plant protection measures are often low, rendering the impact of pests and 635 diseases larger than in high-input regions, (d) yields reported in global databases have often been estimated, and (e) cultivars may be highly adapted to local environmental conditions and consumer preferences.In addition, the fact that EPIC takes only the major stress (out of nutrient deficit, water deficit, adverse temperature, and limited root aeration) on each day of the growing season into account to limit biomass production implies interactions between nutrient input levels and the model's sensitivity to weather-induced stresses.This approach thus lowers 640 the crop's sensitivity to water and temperature stress if nutrient deficits are prevailing.
General conclusions as to which setup factors improve the performance in certain countries are challenging to draw.The setup options represent highly aggregated domains with more than 20 parameters included in the coefficient setup domain alone (Table 1).One setup factor that can clearly be identified to allow for higher performance in Argentina and India are the coefficients of EPIC-IIASA, which result in all 16 combinations in a 645 tscorr with r>0.5 (Figure S3-4).However, in both countries there are also two setups based on the Biogeosciences Discuss., doi:10.5194/bg-2016-527,2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.parameterization of GEPIC in combination with GEPIC's soil handling that allow for r>0.5.Vice versa, the parameterization of EPIC-IIASA causes a worse performance in Brazil if combined with the soil data of EPIC-IIASA, but a high tscorr if combined with the cultivar and soil data of GEPIC.The most dramatic contrast in setups and their performance is found in Indonesia and Ukraine, where those resulting in higher performance in 650 the first cause lower performance in the latter, albeit none of the scenarios yields r>0.5.Identifying country-or region-specific optimal setups will hence require an even wider range of setup permutations, ideally based on more detailed datasets or the inclusion of findings from local and field-scale model studies that allow for reducing the dimensions of such an exercise.Also, the different setups can represent part of the variability in management systems that are present in actual production systems but greatly underrepresented in modelled 655 production systems.
Besides setups and algorithms of the models, also input data (e.g.Ruane et al., in preparation), validation data (Müller et al., 2016), and landuse masks (Porwollik et al., 2016) affect model performance, complicating its evaluation (see Müller et al. 2016).Thus, the focus of this study was not on model performance per se, but rather on identifying differences among EPIC-MFWs that substantially affect model performance.660

Conclusions
The results presented herein highlight the limited comparability of outputs from various GGCMs based on the same, or very similar, sets of biophysical algorithms, but with differences in parameterizations and management assumptions.This has implications as well for (a) the preparation of common input data for future ensemble runs aiming at the identification of differences in model algorithms globally, and (b) the interpretability of 665 GGCM-based studies.On the other hand, the EPIC-based ensemble shows better agreement in terms of mean model bias if certain aspects of model uncertainty are eliminated (growing seasons, fertilizer application levels, and soil nutrient supply), in contrast to a non-EPIC-based ensemble which shows increasing bias.
The differences in model outputs induced by differences in setups indicate that further steps of harmonization among GGCMs should be taken if model algorithms are to be compared globally.However, prioritizing and 670 selecting input data for further harmonization of the whole ensemble is not a straightforward process.None of the MFWs evaluated herein exhibited an outstanding performance in reproducing reported yields that would allow for justifying the selection of its maize cultivar or wheat type distribution.In order to overcome issues related to the wheat type distribution, following phases of the GGCMI project will collect winter and spring wheat simulations separately for regions in which both may be grown in order to allow for ex-post distributions 675 according to wheat type masks.A harmonization of major maize cultivars appears less feasible in the short run due to differing scopes among research groups.Although the cultivar distribution of EPIC-IIASA performs overall better than that of GEPIC, it is as well not optimal in all regions.Producing global cultivar distribution maps may hence require the compilation of a global inventory of representative maize cultivars and their geographic distribution.Soil data will need to be harmonized to avoid differences in nutrient supply in low-input 680 regions from SOM mineralization and especially differences in soil hydrology in regions with low precipitation under rainfed growth conditions.While a compilation of global crop management practices cannot be expected in the short run, management practices regarding the timing of fertilizer application and thresholds for automatic fertilizer and water irrigation need to be harmonized to avoid deviations among GGCMs impairing the interpretability of plant water stress.To address this issue, model runs in GGCMI phase 2 will be performed with prescribed timing and rates of fertilizer application.Plant residue management can differ substantially from farm to farm as it is subject to farm type and socio-economic drivers and may hence rather be addressed by scenario analyses in future GGCM studies.
Further harmonization of GGCMs in terms of subroutines and parameterization, on the other hand, can be expected to become a long-term process and care must be taken to ensure that the harmonized variables are a 690 genuinely improved representation of reality, and not just a more limited sampling of the uncertainty space.
Single parameters and subroutines such as PET estimation methods or microbial turnover processes are often representative for specific climatic and agro-ecologic regions.However, up to now they cannot be attributed spatially based on any available covariates in global datasets.Hence, running a range of MFWs based on the same field-scale model in different parameterizations allows meanwhile to bracket likely yield estimates by 695 covering gaps in the understanding of agro-environmental processes.Further regional and field assessments, in combination with climate and pedologic zoning, will be required to derive reasonable spatially explicit datasets on the optimal PET estimation method, parameterization of microbial activity, and common SOM turn-over rates among others.0.9 0.99 0.8 0.9 0.9 Coefficient allocating root growth 0.5 0.5 0.7 0.5 0.5 N volatilization coefficient 0.005 0.7 0.03 0.005 0.3 1) Annual renewal of soil profile if a static soil profile is used and wind/water erosion is turned on 2) Describes the dependence of curve number (CN) estimation on soil moisture, which can be based on five methods, among them soil moisture gradient with profile depth or calculation of a daily soil moisture index (SMI) 3) Field capacity (FC) and wilting point (WP) can be estimated by 10 different methods -among them Rawl's method used by most MFWs -

925
or be input directly after prior estimation, which is the case for EPIC-IIASA where the Rosetta model had been applied a priori.
4) The auto-fertilizer and irrigation triggers define at which stress level fertilizer or water are being applied.E.g., a value of 0.8 for the autofertilizer trigger implies that fertilizer is applied on a given day if potential biomass production would be limited by >20%.Zero indicates that fertilizer or irrigation water is not applied automatically.
at a spatial resolution of 0.5°x0.5°were provided by the ISI-MIP and GGCMI projects.The climate data are based on temperature and Biogeosciences Discuss., doi:10.5194/bg-2016-527,2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.

Figure
Figure S2-2; TableS2-2).The most distinct differences were found in sub-Saharan Africa, particularly the Biogeosciences Discuss., doi:10.5194/bg-2016-527,2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.India, and Southeast Asia.The lowest differences with CV usually <100% and often <50% are found in mid and high latitudes of both hemispheres, where (a) yields are at moderate or high levels, (b) most models plant the same high-yielding cultivar and (c) the annual temperature curve usually defines a narrow growing season window.Rainfed cultivation results in larger differences among models in (semi-)arid regions of Central and West Asia, the Western USA and North-eastern Brazil.If irrigation water is applied by the MFWs, differences 345 increase in most parts of sub-Saharan Africa and Central India, but decrease in most of North and South America, Central Asia, and Europe.
yield variability indicate that yield estimates in individual years are not only affected by climate but also strongly by additional factors such as parameterization of soil hydrology and evapotranspiration, cultivars, and timing of irrigation water supply.Biogeosciences Discuss., doi:10.5194/bg-2016-527,2016 Manuscript under review for journal Biogeosciences Published: 20 December 2016 c Author(s) 2016.CC-BY 3.0 License.

Figure 1 :
Figure 1: Distributions of maize cultivars in the model frameworks for (a) EPIC-IIASA, (b) EPIC-TAMU, (c) GEPIC and PEPIC, and (d) EPIC-BOKU.Differences in the parameterization of each cultivar are provided in Table S1-1.Numbers in braces (1-4) are used in the text to refer to the cultivars. 960

Figure 5 :Figure 6 :
Figure 5: Global average rainfed maize yields over a 31 year time period for 32 setup combinations based on the EPIC-IIASA and GEPIC setups as described in Table 3. e=EPIC-IIASA, g=GEPIC, Cult=cultivar definition and distribution, SoilD=soil data, SoilP=spin-up and soil handling, Coeff=coefficients, Manage=crop management (indicated by line color).Dashed lines show linear regressions to identify trends.