Improving the representation of anthropogenic CO 2 emissions in climate models : impact of a new parameterization for the Community Earth System Model ( CESM )

ESMs (Earth System Models) are important tools that help scientists understand the complexities of the Earth’s climate. Advances in computing power have permitted the development of increasingly complex ESMs and the introduction of better, more accurate parameterizations of processes that are too complex to be described in detail. One of the least well10 controlled parameterizations involves human activities and their direct impact at local and regional scales. In order to improve the direct representation of human activities and climate, we have developed a simple, scalable approach that we have named the POPEM module (POpulation Parameterization for Earth Models). This module computes monthly fossil fuel emissions at grid point scale using the modeled population projections. This paper shows how integrating POPEM parameterization into the CESM (Community Earth System Model) enhances the realism of global climate modeling, improving this beyond simpler 15 approaches. The results show that it is indeed advantageous to model CO2 emissions and pollutants directly at model grid points rather than using the same mean value globally. A major bonus of this approach is the increased capacity to understand the potential effects of localized pollutant emissions on long-term global climate statistics, thus assisting adaptation and mitigation policies. 20


Introduction
The Earth system is a complex interplay of physical, chemical, and biological processes that interact in nonlinear ways (Ladyman et al., 2013;Lorenz, 1963;Rind, 1999;Williams, 2005).Much effort has been devoted to understanding these complex interactions, and several improvements have been made since the end of the last century.
One of the most important advances in this field has been the use of coupled numerical climate models, dubbed Earth system models or ESMs (Edwards, 2011;Flato, 2011;Schellnhuber, 1999).These models aim to simulate the complex interactions of the atmosphere, ocean, land surface, and cryosphere, together with the carbon and nitrogen cycles (Giorgetta et al., 2013;Hurrell et al., 2013;Martin et al., 2011;Schmidt et al., 2014).
One of the fields most in need of development is the inclusion of co-evolutionary dynamical interactions of the socioeconomic dimension into global models with other Earth system components (Nobre et al., 2010;Robinson et al., 2018;Sarofim and Reilly, 2011).Human activity has become a major driver of change in the Earth system, especially over the past several decades (Alter et al., 2017;Barnett et al., 2008;Crutzen, 2002), and it now dominates the natural system in many different ways (Motesharrei et al., 2016;Ruth A. Navarro et al.: Improving the representation of anthropogenic CO 2 emissions in climate models et al., 2011).However, most global models use basic socioeconomic assumptions about the behavior of societies and are only unidirectionally linked to the biogeophysical part of the Earth system (Müller- Hansen et al., 2017;Smith et al., 2014).The standard way of introducing anthropogenic climate change into ESMs is through representative concentration pathways (RCPs).These are consistent sets of projections involving only radiative forcing components (van Vuuren et al., 2011), but which represent a step forward from the scenario approach of the last decade (Moss et al., 2010;van Vuuren et al., 2014;van Vuuren and Carter, 2014).However, RCPs are not fully integrated socioeconomic parameterizations but rather estimates for describing plausible trajectories of human climate change drivers (Moss et al., 2010;van Vuuren et al., 2012).They provide simplified accounts of human activities and processes from one-way coupled integrated assessment models (IAMs; Müller- Hansen et al., 2017).
The use of RCPs is advantageous because they provide a set of pathways that serve to initialize climate models.However, two major problems remain within this approach.Firstly, human activities are not intrinsically embedded into the ESM, impeding sensitivity studies.Secondly, because of the weak coupling of IAMs, they cannot capture the sometimes counterintuitive bidirectional feedback and nonlinearity between the socioeconomic and natural subsystems (Motesharrei et al., 2016;Ruth et al., 2011).Good examples that illustrate the importance of including such bidirectional feedbacks feature in the HANDY model (Motesharrei et al., 2014) which has been used to analyze the key mechanisms behind societal collapses.
The RCP approach has been used in climate models because of its low computational cost.However, advances in computational resources now allow to parameterize human-Earth processes in a more detailed way, including the inclusion of population dynamics into the modeling, as in the POPEM (POpulation Parameterization for Earth Models) module (Navarro et al., 2017).
One important, but sometimes overlooked, process is the direct regional effect of anthropogenic greenhouse gas (GHG) emissions.Although some GHGs quickly mix in the atmosphere (IPCC, 2014a), their mixing times and lifetimes vary (Archer et al., 2009;Prather, 2007), and localized emissions may produce a transient response in the atmosphere.Given the highly nonlinear character of the processes involved, it is not unreasonable to assume that accounting for geographical variability is significant, and the spatial and time distribution of these emissions may affect global climate (Alter et al., 2017;Grandey et al., 2016;Guo et al., 2013).This hypothesis has seldom been investigated, as most current models treat certain GHG emissions as a homogeneously distributed forcing.Thus, for instance, the most typical CESM (Community Earth System Model) simulations prescribe a CO 2 concentration on the assumption that it is well mixed in the atmosphere (Neale et al., 2012).This paper describes the results of a 50-year simulation with a simple parameterization of fossil fuel CO 2 emissions at model grid-point scale, integrating the POPEM module into the CESM.The aim of this paper is to show that this grid-point scale modeling of anthropogenic CO 2 emissions (and other pollutants) represents an improvement over simpler approaches, and leads to better representation of the geographical variability of precipitation.
The purpose of the new modeling is not only to improve precipitation and temperature estimates but also to help understand the carbon cycle feedback, and evaluate the climate sensitivity of the Earth under alternative GHG emission scenarios.While our focus here is anthropogenic CO 2 emissions, the POPEM parameterization can accommodate other GHGs and human-dependent processes in order to advance CESMs towards a comprehensive fully coupled modeling of anthropogenic dynamics in the global climate.
The paper is organized as follows: in Sect.2, we present the validation of the POPEM stand-alone mode and set the framework for evaluating the impact of POPEM parameterization -its incorporation into the CESM and the testing framework; in Sect.3, we compare the outputs of CONTROL and POPEM runs and see how they compare with observations.In Sect.4, we highlight the importance of the dynamical modeling of anthropogenic emissions at grid-point scale to better represent the socioeconomic parameters in the CESM model and improve precipitation estimates.

The CESM model
The Community Earth System Model (CESM) is a stateof-the-art ESM and probably the most widely used climate model.It was developed and is maintained by the National Center for Atmospheric Research (NCAR), with contributions from external researchers funded by the US Department of Energy, the National Aeronautics and Space Administration (NASA), and the National Science Foundation (Hurrell et al., 2013).CESM is an ESM comprising a system of multigeophysical components, which periodically exchange twodimensional boundary data in the coupler (Craig et al., 2012).It consists of five component models and one central coupler component: the atmosphere model CAM (Community Atmosphere Model; Tilmes et al., 2015); the ocean model POP (Parallel Ocean Program; Kerbyson and Jones, 2005); the land model CLM (Community Land Model; Lawrence et al., 2011); the sea ice model CICE (Community Ice Code; Hunke and Lipscomb, 2008); and the ice sheet model CISM (Community Ice Sheet Model; Lipscomb et al., 2013).
A major advantage of CESM over other ESMs is its availability.Some climate models are developed by scientific groups and access to the source code is limited.The CESM source code is free and available to download from the NCAR website.This approach helps improve the model by setting up a framework for collaborative research and makes the model fully auditable.CESM is a good example of a "full confidence level" model, after Tapiador et al. (2017), where many "avatars" of the code are routinely run in several independent research centers, and there is an entire community improving the model and reporting on issues and results.However, the model is not immune to bias.One important shortcoming is the poor representation of precipitation in terms of spatial structure, intensity, duration, and frequency (Dai, 2006;Tapiador et al., 2018;Trenberth et al., 2015Trenberth et al., , 2017)).Another major bias is the anomalous warm surface temperature in coastal upwelling regions (Davey et al., 2002;Justin Small et al., 2015;Richter, 2015).

POPEM parameterization model overview
The POPEM module is a demographic projection model coded in FORTRAN that is intended to estimate monthly fossil fuel CO 2 emissions at model grid-point scale using population as the input.Due to a lack of actual GHG measurements at appropriate spatial and temporal scales, it is necessary to use a proxy.For this, POPEM employs population, the evolution of which is modeled using external parameters that feed the module.The idea of using population as proxy is not new, and population density has previously been used to downscale national CO 2 emissions (Andres et al., 1996(Andres et al., , 2016)).However, these inventories were not dynamical but instead tied to historical data so it is not possible to use them either to estimate future changes in emissions or coupled with other components of the model.This change represents an important advance in the way emissions are computed.Thus, POPEM uses a bottom-up approach, where emissions are calculated at cell level on the basis of population projections, while global inventories use a top-down approach, which is less flexible when coupled with other components of the ESM.
The demographic/emissions module presented here is an updated version of the demographic module explained in Navarro et al. (2017).The differences between the versions are minimal.They involve better approximation of emissions in highly polluting regions with poor population data, such as China; a better estimate for coastal zones and country limits; and a change in the model time step for more efficient coupling with CESM.The inclusion of these changes results in more accurate emission estimates when compared with inventories than the previous version did.However, the model is not immune to bias.The most important limit is the degradation of the model outputs when there is increased spatial resolution -resolution of 0.25 • and higher.
Detailed information on POPEM and its validation in the demographic realm can be found in Navarro et al. (2017).In short, from an initial condition, the routine computes the population for each model grid point in a 2-D matrix and then calculates fossil fuel CO 2 emissions using per capita emission rates by nations.The process is repeated for each time step (e.g., annually) throughout the simulation period.
As seen in Fig. 1, POPEM stores gridded emission data in a 3-D array (time, latitude, and longitude) to be used by the modified version of the co2_cycle module.This module reads emission data and passes this to the atm_comp_mct, which calculates the total amount of CO 2 emissions from different sources (land, ocean, and fossil fuel).

POPEM trend verification
Prior to coupling POPEM with CESM, we performed several tests to evaluate its ability to reproduce historical population trends and CO 2 emissions.To do this, we ran the module in stand-alone mode.In a first test, we ran a short simulation  and compared the emission data with a standard emissions inventory (CDIAC).In a second test, POPEM was run for 70 years  and population estimates were validated against the UN (United Nations) population statistics database for those years when data were available.
As shown in Fig. 2, POPEM is capable of satisfactorily simulating the observed population.Comparison with UN data shows good agreement.However, POPEM presents slight differences from the reference data in some regions.Several of these discrepancies can be explained by the initial model conditions; POPEM uses the same age distribution inside each grid cell to initiate the model (only for the first time step).This distribution is based on the global average age structure.Consequently, the model overestimates the popu-   Figure 1.Conceptual scheme of the POPEM module coupled with the CAM5 atmosphere module.POPEM requires three input data sets to compute emissions (black dashed rectangles): initial population distribution; demographic parameters (age structure, death, and birth rates); and per capita emission rates by country.POPEM provides a 3-D array (time, latitude, longitude) with emissions that are read by the co2_cycle module and passed to the atm_comp_mct module which computes the total amount of CO 2 in the atmosphere.lation in those regions with a more elderly age structure, i.e., Europe and North America, and underestimates areas with younger populations, i.e., Latin America and Asia.
These disparities in population counts have a diverse effect on the outputs in terms of GHG emissions.Thus, for example, the bias in Europe seems to be more important than the bias in Latin America and Oceania.Two principal reasons could explain this: population size, as Europe has a larger population than Oceania, so there is greater bias in the CO 2 emissions estimation; and the per capita emissions rate, as Latin American countries have lower per capita emission rates than European nations.
It is worth noting here that the POPEM outputs in Fig. 2 are clearly nonlinear and thus not trivially derived from simply extrapolating population.The North American estimate of CO 2 emissions (second row from the bottom) clearly shows the added value introduced by the model.The first row compares global data, the second to seventh rows compare regional data (Africa, Europe, Latin America, North America, and Oceania).In (a), the red line shows the estimates given using POPEM and blue indicates UN estimates.Values are given in billions of people.In (b), the red line shows the estimates given using POPEM and the black indicates CDIAC estimates.Units are given in million metric tons.
Figure 3 shows how POPEM distributes CO 2 emissions for different years in the recent past.In 1950, the majority of emissions tended to be concentrated in the USA and Europe, while in 2000, China, the USA, and India were the most polluting countries.This is consistent with the literature: POPEM's estimates generally agree with the emissions maps for the recent past (Andres et al., 1996;Boden et al., 2017;Oda et al., 2018;Rayner et al., 2010), as well as with regional studies on CO 2 emissions (Gately et al., 2013;Gurney et al., 2009).The regionalized distribution of emissions depicted in Fig. 3 represents a vast improvement over the standard procedure of using globally averaged emissions.Even accounting for rapid mixing of GHGs, transient effects are likely to appear given the hemispheric contrast and regional differences in the emissions.The differences in Asia are illustrative of the economic changes in the recent past and the exponential pace of industrialization in that region.

CESM experimental setup
The CESM used in this work is based on version 1.2.2 (http://www.cesm.ucar.edu/models/,last access: 10 February 2018).This set includes active components for the atmosphere, land, ocean, and sea ice, all coupled by a flux coupler.The latest atmospheric module CAM5 (Neale et al., 2012) is used to introduce more accurate modeling of atmospheric physics.Additionally, the carbon cycle module is included in CESM's atmosphere, land, and ocean components (Lindsay et al., 2014).
We ran an experiment at 1.9 • of spatial resolution for the period 1950-2000.Two simulations were performed to analyze the effects of the regionalized emissions (Fig. 3) on the CESM.Our control case used homogeneous CO 2 concentration parameters (standard procedure in ESMs), while the POPEM case used geographically distributed CO 2 emission data.In the latter, the POPEM module was coupled with the atmospheric CO 2 flux routine to provide monthly gridded CO 2 emissions.The gridded data were used at each time step by the atmospheric routine.Apart from this change, both simulations were identical in order to identify the effects (if any) of the POPEM parameterization.

Validation data 2.4.1 GPCP data set
Precipitation is one of the key elements for balancing the energy budget, and one of the most challenging aspects of climate modeling.Hence, high-quality estimates of precipitation distribution, amount, and intensity are essential (Hou et al., 2014;Kidd et al., 2017;Xie and Arkin, 1997).While there are many sources of precipitation data to be used as a reference (see Tapiador et al., 2012, for a review), only a few qualify as "full confidence level validation data" (Tapiador et al., 2017).
The Global Precipitation Climatology Project (GPCP; Adler et al., 2016) has several products suitable for validating climate models.GPCP-Monthly is one of the most popular precipitation data sets for climate variability studies.It combines data from rain gauge stations and satellite observations to estimate monthly rainfall on a 2.5 • global grid from 1979 to the present.The careful combination of satellite-based rainfall estimates results in the most complete analysis of rainfall available to date over the global oceans, and adds necessary spatial detail to rainfall analyses over land.Due to its relevance and global coverage, it has been widely used for validating precipitation in climate models (Li and Xie, 2014;Pincus et al., 2008;Stanfield et al., 2016;Tapiador, 2010).

CRU data set
Global surface temperature data sets are an essential resource for monitoring and understanding climate variability and climate change.One of the most commonly used data sets is produced by The Climate Research Unit at the University of East Anglia (CRU).This group produces a high-resolution gridded climate data set for land-only areas, the Climate Research Unit Time-series (CRUTS; Harris et al., 2014).CRUTS contains monthly time series of 10 climate variables, including surface temperature.The data set is derived from monthly observations at meteorological stations.Station anomalies are interpolated into 0.5  (New et al., 1999(New et al., , 2000)).It is commonly used in the validation of climate models because of its confidence levels, together with temporal and spatial coverage, and the fact it compiles station data from multiple variables from numerous data sources into a consistent format (Christensen and Boberg, 2012;Hao et al., 2013;Liu et al., 2014;Nasrollahi et al., 2015).

GISTEMP data set
NASA's GISTEMP (GISS Surface Temperature Analysis) is a global surface temperature change data set (Hansen and Lebedeff, 1987; see Hansen et al., 2010, for an updated version).It combines land and ocean surface temperatures to create monthly temperature anomalies at 2 • × 2 • degrees of spatial resolution.The use of anomalies reduces the estimation error in those places with incomplete spatial and temporal coverage (Hansen and Lebedeff, 1987).The anomalies are calculated over a fixed base period  that makes the anomalies consistent over long periods of time.
The first version was originally conceived only for land areas (Hansen and Lebedeff, 1987) but in 1996 marine surface temperatures were added (Hansen et al., 1996).The updated version of GISTEMP includes satellite-observed night lights to identify stations located in extreme darkness and adjust temperature trends of urban stations for non-climatic factors (Hansen et al., 2010).Just like CRUTS, GISTEMP is commonly used to validate climate models because of its coverage and confidence levels (Baker and Taylor, 2016;Brown et al., 2015;Neely et al., 2016;Peng-Fei et al., 2015).

Comparison between the CONTROL and POPEM runs
It is worth stressing that a parameterization which performs well when tested for the variable it models does not necessarily translate into an overall improvement of the other variables in the model.An accepted practice in climate modeling is to tune ESMs by adjusting some parameters to achieve a better agreement with observations (Hourdin et al., 2017;Mauritsen et al., 2012).These adjustments to specific targets may, however, decrease the model's overall performance (Hourdin et al., 2017), and give poor scores for variables other than those tuned.Thus, for example, if a model is biased with respect to aerosol concentrations or humidity, then improved parameterization of cloud formation may worsen the performance of the model with regard to precipitation (Baumberger et al., 2017).This mismatch can be caused by model over-specification, or over-tuning.
The first step in evaluating the new parameterization is to compare the outputs with a control simulation to make sure the new addition does not negatively interact with the dynamical core or spoil the contributions of the rest of the parameterizations.Figure 4 shows that this is not the case with the POPEM parameterization, which does not negatively affect the outputs of precipitation and temperature.Rather, both variables are now closer to the observed data than they were in the control run, especially in terms of reducing the double ITCZ (Intertropical Convergence Zone), which artificially features in global models (Mechoso et al., 1995; for a recent analysis of double ITCZ in CMIP5 models see Oueslati and Bellon, 2015).
Figure 4a shows that there is just a slight discrepancy in the absolute difference in rainfall between the GPCP and CESM simulations (the first and the third quartiles of the distribution remain between ±0.4 mm day −1 ).Grid-point to grid-point comparison between the model and GPCP indicates the ability of CESM to reproduce the spatial distribution of precipitation.In both simulations, the CESM exhibits a good correlation coefficient (0.72 R 2 ) compared with the reference data (Fig. 4b).The results are even better for temperature (0.88 R 2 ; Fig. 4d).
Direct comparison of aggregated data is a standard procedure for gauging model abilities.Figure 5 compares two latitude-time graphs for precipitation (Fig. 5a) and surface temperature (Fig. 5b), both for the CONTROL case and for the new POPEM parameterization.
It is clear from Figs. 5a and 6a that POPEM does alter the spatial pattern of precipitation and exerts a definite effect on the climate pattern, as the module reduces the otherwise exaggerated ITCZ precipitation in the Southern Hemisphere reported by several authors (Hwang and Frierson, 2013;Li and Xie, 2014).
Disparities in temperature between the CONTROL and POPEM runs are apparent at high latitudes.In this case, POPEM produces lower temperatures at both poles, a result which deserves further attention (Figs.5b and 6b).
There are also important differences in precipitation in the 30 • N-30 • S band.Here POPEM reduces model bias, especially in the Southern Hemisphere and on the Tibetan Plateau (see Sect. 3.2 for more details).On the other hand, POPEM departs from the control simulation in the Asia Pacific region between 10 • N-10 • S.This result reinforces the double ITCZ bias in this area.
These results show that the POPEM parameterization generally agrees with historical data for population, and also compares well with the control simulation in the sense of addressing some of the known biases in precipitation and temperature, offering a more detailed version of CO 2 emissions at a relatively cheap computational cost.As discussed above, the CONTROL run uses global concentration values to include CO 2 on the assumption that it is well mixed in the atmosphere (Neale et al., 2012).This assumption reduces the computational burden of the simulation but does not al- low for precise emissions modeling in the future.This is an important aspect for regionalized emission scenarios, since even if the new parameterization is not significantly better than the old approach (but no worse), it is desirable as it allows for sensitivity analyses, such as evaluating the effects of the US leaving the Paris Agreement.Potential applications of POPEM include not only sensitivity analyses of local CO 2 emissions policies but also the added feature of performing tests for "what if" scenarios.One interesting example would be the climate response under the hypothesis that China and India -the most populated countries in the world -reach US CO 2 per capita emission rates.Another "what if" scenario would be the climate response of an increasingly urbanized world.In both cases, POPEM provides a flexible framework for testing the alternative hypotheses.
The realism of the ESM will be enhanced with a fully coupled system.Such a fully fledged ESM will include bidirectional feedback between POPEM and CESM to evaluate the effects of climate change on population dynamics and emissions.

Validation against observational data sets
Once it has been verified that the new parameterization does not worsen the modeling, the next step in evaluating the performances is comparing the simulation outputs for both the CONTROL run and the POPEM module using actual observational data.Direct comparisons with historical data can help show whether or not a climate model correctly represents the climate of the past.However, although observational measurements are often considered the ground truth to validate models against, it is important to be aware that measurements have their own uncertainties (Tapiador et al., 2017).
Figure 7 shows a comparison of CESM precipitation simulations for the period 1980-2000 using the GPCP.It is apparent that there is an overall consensus, even though there are differences.Despite these known biases, the model agrees with the observations on the major features of global precipitation.
The improvements in parameterizing emissions become clearer if we focus on specific regions.For the El Niño-4 area, there are statistically significant differences (at the 0.05 significance level) between both the CONTROL run and the POPEM modeling when compared with the refer-  ence data.This observation illustrates the limitations of the modeling and the need of advances in the parameterizations.However, for this area the correlation (R 2 ) between POPEM and GPCP is slightly better than CONTROL and GPCP (0.706 R 2 versus 0.692 R 2 ).The real added value, however, is not in a better estimation of the totals but in the ability of POPEM to better capture the structure of the precipitation.Figure 8 shows the histograms of mean precipitation in the El Niño-4 area using the POPEM parameterization (Fig. 8a), the standard forcing approach (CONTROL, Fig. 8b), and the reference GPCP estimates (Fig. 8c).While the CONTROL simulation severely overestimates the low end of the distribution, POPEM gives a more realistic value.This result is not apparent in the otherwise improved correlation of POPEM, and is also buried in the box plots.
El Niño-4 is important because it presents a lower variance in the SST (sea surface temperature) than any other of the El Niño areas, playing a key role in identifying El Niño Modoki events (Ashok et al., 2007;Ashok and Yamagata, 2009;Yeh et al., 2009).The consequences of such events are severe disruptions in human activities due to the increased risk of droughts, heat waves, poor air quality, and wildfires (McPhaden et al., 2006).Thus, precise modeling of the processes in this sector of the Pacific is extremely important.
Another important benefit of POPEM is the reduction of the double ITCZ bias in the Southern Hemisphere.Although a small change can be inferred from Fig. 7a and b, the improvement is buried in the annual mean precipitation maps.
Figure 9a shows that the POPEM results are closer to observations of the intra-annual variability in precipitation, especially for the driest months (June-October).
The figure also shows slight improvements for two other typical biases seen in CESM, namely the excess precipitation in the Tibetan Plateau (Chen and Frauenfeld, 2014;Su et al., 2013;Fig. 9c) and the bias in some areas affected by the Asian-Australian monsoon (AAM), such as the top end of Australia (Meehl and Arblaster, 1998;Meehl et al., 2012;Fig. 9b).
The results for the El Niño-4 area show that detailed, gridpoint emissions of GHGs improve the quantification of precipitation in dry areas, in agreement with our hypothesis about the benefits of locally distributed versus global mean forcings.Also, the double ITCZ example shows that the transient effects of regionalized GHG emissions may even translate into (long) 50-year climatologies, meaning there is room for improvement in the "rapidly mixing, well-mixed gases" forcing approach.
Figure 10 compares the annual mean temperatures for the period 1950-2000.CESM simulations show a significant bias in high latitudes of the Northern Hemisphere (cf.Fig. 10a and b).In these areas, the model produces colder temperatures than those registered in the CRUTS reference data but this is also an issue in the CONTROL run.This deviation is also apparent in Fig. 4b, where negative values lie away from the idealized regression line, and indicate further improvement of the CESM.
The bias is also reproduced when compared with temperature anomalies for a specific region.Thus, for instance, CESM gives poor scores in the Barents Sea area (Fig. 11a) while POPEM obtains better results for the Bering Sea, especially in the Russian part (Fig. 11b).Here, POPEM gives more realistic values for the period 1970-1998 but, even with the improvement, the model still overestimates the temperature anomaly.
If we focus on global temperature anomalies, CESM simulations are able to reproduce the progressive increase in the temperature anomaly (Fig. 12a).However, the CONTROL case simulates a sharp drop at the end of the period (1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999), while POPEM portrays this change as smooth, in agreement with the observations.The differences between CONTROL and POPEM are better demonstrated when comparing land and ocean separately (Fig. 12b and c).While the temperature anomalies for land are quite similar in both cases, POPEM provides a better representation of the ocean tendency from 1992 onwards, and that translates to an overall improvement (Fig. 12a).

Validation against ESPI and ONI indices
The El Niño-Southern Oscillation (ENSO) is the most dominant inter-annual climate variation in the tropics.It occurs when seasonally averaged SST anomalies in the eastern Pacific Ocean exceed a given threshold and cause a shift in the 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 -0.9 -0.6 -0.3 0 0.3 0.6 0.9      atmospheric circulation (Trenberth, 1997).Historically, the definition of ENSO does not include precipitation because of the limitations of stations (Ropelewski and Halpert, 1987), but recent work with satellites has confirmed that this phenomenon is a major driver of global precipitation variability (Haddad et al., 2004).A major advantage of satellite-derived precipitation indices over more conventional ones is the description of the strength and position of the Walker circulation (Curtis and Adler, 2000).Under that assumption, Curtis and Adler (2000) derived three satellite-based precipitation indices: the ENSO precipitation index (ESPI), El Niño index (EI), and La Niña index (LI).Precipitation anomalies are averaged over areas of the equatorial Pacific and Maritime Continent -where the strongest precipitation anomalies associated with ENSO are found -to construct differences or basin-wide gradients (Curtis, 2008).
Figure 13 shows a comparison of GPCP, CONTROL, and POPEM for the ESPI, EI, and LI indices.
Unfortunately, CONTROL and POPEM cases have difficulty simulating the precipitation patterns associated with ENSO. Figure 13 shows that bias increases in 1982-1983 and 1997-1998 El Niño years.The same bias emerges when comparing the EI and LI indices.In that case, the CESM model produces stronger El Niño/La Niña events than the observed data.Consequently, we can consider that CESM is unable to obtain a precise estimate of precipitation patterns, suggesting that current climate models are far from generating realistic simulations of the precipitation field (Dai, 2006)  Another widely used ENSO index is the Oceanic Niño Index (hereafter ONI).ONI was developed by the NOAA Climate Prediction Center (CPC) as the principal means for monitoring, assessing, and predicting ENSO (Kousky and Higgins, 2007).This index is defined as 3-month runningmean values of SST departures from the average in the Niño-3.4region.It is computed from a set of homogeneous historical SST analyses (Kousky and Higgins, 2007;Smith et al., 2003).
Figure 14 compares the ONI index for CPC, POPEM, and CONTROL cases.It is clear from the figure that POPEM produces a more realistic representation of the ENSO, especially 1979ENSO, especially 1983ENSO, especially 1987ENSO, especially 1991ENSO, especially 1995ENSO, especially 1999  if we focus on the 1992-1999 period.POPEM also obtains better results than CONTROL in the number of simulated El Niño events (see Table 1).The improvement is also noticeable in the intensity.The CONTROL case exhibits an overly strong ENSO -a common bias in CESM (Tang et al., 2016) -but POPEM reduces this bias (0.22

Conclusions and future work
Like all models, climate models are simplified versions of the real world and therefore do not include the full complexity of the Earth system.Due to certain limitations, e.g., computational resources or spatial and temporal resolution, climate models have to make assumptions and resort to parameterizations.
One important simplification is to use prescribed forcings instead of dynamically modeling GHG emissions.However, precise modeling of anthropogenic CO 2 emissions is important for climate change research as it allows sensitivity analyses to be performed.
Here we present a new module of gridded CO 2 emissions that is coupled with CESM.The module, denominated POPEM, computes anthropogenic CO 2 emissions by using population estimates as a proxy for disaggregating emissions beyond the national level.POPEM makes CESM use dynamical emission data instead of fixed concentration parameters.
In terms of population and emissions, the module compares well when validated with data.Thus, POPEM's estimates for the 1950-2000 period are in general agreement with population and emission inventories from the recent past.In spite of the more realistic depiction of the actual emissions (Fig. 3), issues persist.The performance of the model can be further improved in places where popula-tion projections are difficult to model.For instance, POPEM tends to underestimate emissions on the west coast of the United States and the Anatolian Plateau, and overestimates emissions in China and Japan.
When the POPEM module is coupled with CESM to generate climatologies, the ability to successfully model precipitation and surface temperature is preserved.Moreover, the results of 50-year simulations show that the dynamical modeling of emissions produced by POPEM results in slight but noticeable differences in the resultant precipitation regime and surface temperature.Thus, dynamically modeling the emissions alters the ITCZ by reducing precipitation in the Southern Hemisphere and increasing it in the Northern Hemisphere.For particularly interesting areas, such as the El Niño-4 region, the POPEM outperforms the traditional approach.
Further work will be devoted to improving the modeling of those areas and hopefully minimizing some of the original biases of the CESM model.These include the emergence of a double ITCZ in CESM simulations, which is a common bias for most climate models (Oueslati and Bellon, 2015), as well as SST simulated by climate models, which are generally too low in the Northern Hemisphere and too high in the Southern Hemisphere (Wang et al., 2014).
Current applications of the parameterization include evaluating the effects of changes in regional policies, and a better understanding of the carbon cycle (Friedlingstein et al., 2006).Future work will be devoted to evaluating the climate response to alternative anthropogenic CO 2 emissions, to coupling POPEM with the newest version of CESM (CESM2; Joel, 2018), to fully coupling human-Earth subsystems, to increasing the spatial resolution of the simulations, and to refining the spatial and temporal distribution of emission estimates.
Although the version of POPEM presented here is already functional, this work is intended to be just the first step in fully coupling socioeconomic dynamics with ESMs.This will include bidirectional feedbacks between human and Earth systems and the simulation of societal processes based on the internal dynamics of the model instead of using external sources to make the projections.Only within a coupled global human-Earth system framework can we produce more realistic representations of the Earth system capturing much of the important feedbacks that are missing from current models (Motesharrei et al., 2016).The success of this approach will depend on the ability of scientists from different research fields to work in an interdisciplinary framework of continuous collaboration.

1048A.
Navarro et al.:  Improving the representation of anthropogenic CO 2 emissions in climate models

Figure 2 .
Figure 2. Comparison of the population estimates for the years 1950-2020 (a) and the historical CO 2 emission estimates for the years 1950-2012 (b).The first row compares global data, the second to seventh rows compare regional data (Africa, Europe, Latin America, North America, and Oceania).In (a), the red line shows the estimates given using POPEM and blue indicates UN estimates.Values are given in billions of people.In (b), the red line shows the estimates given using POPEM and the black indicates CDIAC estimates.Units are given in million metric tons.
Figure 3. POPEM CO 2 emission estimates for 1950, 1980, and 2000.POPEM produces a gridded representation of anthropogenic CO 2 emissions using population dynamics and country per capita emissions derived from the CDIAC database.Values are given in millions of metric tons per year.

Figure 4 .
Figure 4. Box plots of CESM simulation bias for precipitation (a) and temperature (c).(b) Scatter plots comparing the annual mean precipitation (1980-2000) at every grid point for GPCP and CESM simulations (POPEM and CONTROL).(d) Scatter plots comparing the annual mean temperature at every grid point for CRU and CESM simulations (POPEM and CONTROL).Units are in mm day −1 (precipitation) and in • C (temperature).

Figure 5 .
Figure 5. Latitude vs. time plots for precipitation (a) and surface temperature (b).For absolute difference graphs, blue represents higher values in POPEM and red represents higher values in the CONTROL.Units are in mm day −1 for precipitation and in • C for temperature.
Figure 6.A comparison of global annual mean precipitation (1950-2000) for the CONTROL and POPEM (a).(b) is a comparison of annual mean surface temperatures.The maps in the right-hand column show the absolute differences between the simulations (CONTROL minus POPEM).In these, blues represent higher values in POPEM and reds represent higher values in the CONTROL.

Figure 7 .
Figure 7.A comparison of the global annual mean precipitation (1980-2000) as simulated by the CESM (POPEM and CON-TROL) model and GPCP observational database.(a) Global annual mean precipitation maps for GPCP, POPEM, and CONTROL.(b) Absolute difference maps.Units are in mm day −1 .

Figure 9 .
Figure 9. Monthly precipitation (1980-1999) based on GPCP, CONTROL, and POPEM for three of the regions with important biases in CESM.(a) Shows precipitation for the area affected by the double-ITCZ bias in the Southern Hemisphere (20 • S-0 , 80 • E-100 • W); (b) for the top end of Australia (30-10 • S, 128-140 • E); and (c) for the Tibetan Plateau (22-32 • N, 78-92 • W).The black line represents observations (GPCP), the blue line is the CONTROL case, and the red line is the POPEM case.Units are in mm day −1 .The arrow indicates the improvement of the POPEM model.

Figure 10 .
Figure 10.A comparison of the annual mean temperature (1950-2000) as simulated by the CESM model (POPEM and CONTROL) and CRU observational database.(a) Global annual mean temperature maps for CRU, POPEM, and CONTROL.(b) Absolute difference maps.Units are in • C.

Figure 12 .
Figure 12.A comparison of the global annual mean surface temperature anomaly between GISTEMP, CONTROL, and POPEM from 1950 to 1999.(a) Global, (b) land, and (c) ocean.The black line represents observational data (GISTEMP), the blue line is the CONTROL case, and the red is the POPEM case.Anomaly was referenced to 1951-1980 period.

Figure 14 .
Figure 14.Comparison of the Oceanic El Niño Index (ONI) for CPC (a), POPEM (b), and CONTROL (c) cases.El Niño and La Niña are defined according to Kousky and Higgins (2007): 3-month running mean with anomalies greater than +0.5 • C (or −0.5 • C) for at least 5 consecutive months in the NIÑO3.4region.The base period for computing SST departures is 1971-1999.

Table 1 .
Comparison of the ONI index for the period 1950-1999.The table compares the ability of the models to reproduce the number, strength, and duration of El Niño events.The number of months that CPC and CESM agree on El Niño. 2 Disagreement defined as the number of months where CPC and CESM obtain opposite results. 3Intensity: (|CESM ONI| − |CPC ONI|)/number of cases (units in • C). 4 Mean duration of El Niño event (in months).