A comprehensive view on climate change: coupling of earth system and integrated assessment models

Climate change is a complex and comprehensive process that can only be understood on the basis of the combined insights from various scientific disciplines. Natural scientists contribute to an improved understanding by looking at issues like the global energy balance, the carbon cycle and changes in atmospheric composition. At the same time, economists, social scientists and engineers provide insights into the drivers of anthropogenic climate change and the options for adaptation and mitigation, and yet other scientists, including geographers and biologists, study the impacts of climate change. Along these lines, it is possible to identify three research communities, sometimes referred to as the 'earth system' (ES), 'vulnerability, impact, and adaptation assessment' (VIA), and 'integrated assessment' (IA) communities⁷. Clearly, important linkages exist between these different areas of research, and in the past the three research communities have already exchanged information on a regular basis. The ES community, for instance, uses the outcomes of IA models in the form of emission and land use scenarios as input to their models (e.g. the IS92 scenarios [1, 2], the SRES scenarios [3] and more recently the representative concentration pathways (RCPs) [4, 5]). Both the IA and VIA communities, in their turn, use climate model output for assessing climate change impacts or socio-economic responses. Many IA modelers have also used simplified climate models, such as the MAGICC model [6, 7], which are calibrated to state-of-the-art ES models.

In recent years the need for integration of information has become stronger [4, 8]. Moss et al [4], for instance, explicitly mention the need for a stronger collaboration between the research fields to better account of the possible feedbacks between human systems and the earth system on the global scale. Papers have also shown how climate impacts may influence macro-economic development [9, 10], or how (anthropogenic) land use and land cover change can interact with climate [11–16] and air pollution [17]. Also from a policy-making perspective, integration of knowledge is required. The focus towards more integration can also be seen in current research trends. The ES community, for instance, has extended its focus from a primary interest in the climate system itself to related topics such as the carbon cycle, land use, hydrology and even urban systems. At the same time, the IA and VIA communities are also looking into these same topics, partly as part of joint scenario development [18, 19]. As result, the research agendas of these communities are starting to overlap and closer collaboration becomes increasingly useful. Examples of advanced cooperation can already be found. Voldoire et al [11] included the IA model IMAGE in the ES model CNRM-CM3 to look into the interaction of land use change and climate. Bahn et al [20] coupled the IA model MERGE to the intermediate complexity ES model C-Goldstein. Hibbard et al [21] describe the research priorities in the field of land use modeling from a multidisciplinary perspective. Van Vuuren et al [5] describe how the new RCP scenarios have been constructed in a cooperation process involving researchers from the ES and IA communities.

In light of these trends, in this letter we focus on the integration of IA and ES modeling activities in order to improve the understanding of important linkages between human development and environmental change. We discuss the different forms of linkages between IA and ES models, which each have strengths and weaknesses (see further in the letter). It is therefore useful to consider which forms of integration are most suited for different purposes. In the letter, we discuss different research areas where IA and ES models could be coupled and relate these to the various options for information exchange.

In both IA and ES research, models are used as tools to quantitatively explore the linkages between various relevant issues and possible future changes. Partly because ES models can build upon more stable 'natural laws', models in the ES field tend to be somewhat more complex than those in the IA field. The cooperation between the IA and ES model communities can take many forms ranging from activities that stimulate information exchange, dedicated joint research programs, data collection or even shared model components. In figure 1, we have schematically indicated the current research areas of the two communities. We use the figure to start describing possible linkages between the communities such as for emissions, land use and climate parameters. Some of these linkages mostly work in one direction. Other linkages, however, lead to feedbacks that either dampen or strengthen the original signal. While many linkages and feedbacks exists within the human and natural earth systems, they also occur between these systems. Obviously, in models, these can only be accounted for if 'code' exists that describes them. Unfortunately, this is not an easy thing to do as it often involves rather 'weak knowledge' that does not automatically translate well into formal equations [22]. For instance, most IA models do not include an impact of climate change on the energy demand for heating and cooling or an impact of extreme droughts on crop yield, based on the fact that knowledge is assumed to be too uncertain. For improving the IA–ES coupling, a critical condition is therefore also to improve the individual model systems that represent the possible responses to either IA or ES information. Here, making some first basic assumptions about how these components could interact, e.g. using simple linear approximations, could provide insight into the potential relevance of interactions justifying or denying more detailed follow-up studies with the coupled approach.

Zoom In Zoom Out Reset image size

Focusing specifically on the information exchange between IA and ES models (the horizontal arrows in figure 1), we can recognize four different ways in which the interaction can be organized.

• (A)  
  Information can be exchanged in one direction from ES to IA models and/or vice versa (as is currently mostly done). Such interaction can be further refined in the spatial, temporal or variable dimension (to take account of more detailed information available within each community).
• (B)  
  The representation of the Earth system within IA models can be further improved by advancing the climate, vegetation, carbon cycle and atmospheric chemistry representation within the IA models.
• (C)  
  The representation of societal elements within ES models can be further improved by advancing, for instance, the modeling of agricultural and water management options within the ES models.
• (D)  
  IA and ES models can be fully coupled online, allowing for more or less instantaneous two-way interactions.

These four different interactions also represent increasing levels of complexity (from A to D). We briefly discuss the advantages and disadvantages of each of these approaches (see table 1 for a summary). Spatial and temporal scales also play a critical role here. While IA models tend to describe socio-economic processes at the regional scale, ES models mostly use a more detailed geographical grid. A similar situation holds with respect to time, where socio-economic models often tend to focus on longer time scales (e.g. year) than ES models (parts of a day). If necessary, the difference in scale can be accounted for using downscaling techniques [23, 24], but this may be complex especially for time-related issues. Another scale issue is that also regional IA and ES models exist. While some of the arguments might also apply for these models, regional modeling systems may introduce additional constraints as well.

The simplest form of information exchange is a one-way exchange of information (A). Successful examples of this form include the ES model simulations on the basis of IA scenarios (emissions, land use) and the extensive use of climate output (temperature and precipitation) to assess water scarcity, yields or health impacts in IA and VIA work (even within IA 'model frameworks' one-way coupling is sometimes used [25, 26]). An important advantage of the method is that the various research communities can continue to work with existing tools and terminology and pursue their own research strategy. The interaction between the research communities occurs via a limited number of 'pipes' that can be clearly defined, allowing for a transparent information exchange (soft coupling). The method also allows for a high degree of flexibility: ES models can use the output of different IA models, and IA models can use the output of different ES models. This allows a mapping of various sources of uncertainty in a transparent way and without the constraints related to some of the more complex methods of interaction (such as limited flexibility due to calibration). A disadvantage is that the method can only capture feedbacks through iterations involving potentially large numbers of model simulations to reach an 'equilibrium' solution. This is not a practical solution given the computational time involved in ES model simulations. The method will thus often lead to some inconsistencies (e.g. crops might be grown in the IA simulation under conditions unsuitable for these crops in the ES simulation). In fact, identifying these inconsistencies might be an important reason for type A coupling (and form the basis for more integrated analysis). Identification and quantification of the consequences of these inconsistencies should be an important feature of type A (and B/C) coupling exercises.

The second approach (B) builds further on the idea that IA models have already been designed as a tool for integration: they include a representation of both the human and natural subsystems. This approach therefore aims at strengthening the interaction by improving the natural system representation within IA models. Often, IA models only calculate changes in global mean temperature, and correlate impacts directly to this parameter (if calculated at all). A more advanced method is to downscale the outcomes of a relatively simple climate model using the patterns retrieved from sophisticated modeling systems or observations. Evidence so far suggests that pattern scaling works well for temperature, but less well for other variables such as precipitation [27–29]. While it is even more difficult to capture variability, e.g. temperature and precipitation extremes, some methods exist and the results might not necessarily be worse than the skill of ES models. Approaches have been designed to correct patterns of temperature increase for regional sulfur emissions [30]; similar methods could possibly be developed to capture the spatial impact of albedo change, tropospheric ozone and other aerosol components (processes that are generally not yet included in IA models). Within this approach, there is ample room to represent uncertainty, as the simplified climate models included in IA models can be calibrated to reflect the range of outcomes of different ES models [31]. There is also a relatively high level of flexibility in designing model experiments focused on the specific research question. The success of this pathway depends on whether processes currently included in ES models can be simplified, by either a simpler process model, statistical modeling or by meta-modeling such as the development of emulators or improved pattern scaling techniques. Obviously, the model design will be limited in detail in its treatment of physical processes.

Approach C builds on the observation that ES models are now beginning to include components traditionally considered part of human systems. For instance, representations of urban land cover have been included in earth system models in order to calculate urban heat island effects or the impact of albedo increases in urban areas [32, 33]. In addition, vegetation models included in ES models are beginning to incorporate not just natural vegetation, but also various crop types relevant to the representation of managed land. The incorporation of irrigation and the ability to specify nitrogen-based fertilizer application further extends the human management aspects covered by ES [34]. These extensions provide an opportunity to test the implications of alternative human decisions in ES models. For example, sensitivity analyses exploring the magnitude and extent of a particular aspect of land use could be carried out without requiring new inputs from IA models [13]. The increasing computing facilities allow to apply these ES models at increasingly higher spatial resolutions. Obviously, this method will be more complete in its coverage of physical processes than socio-economic processes. The size of ES model runs may limit the treatment of uncertainty.

The fourth approach (D) is full online coupling of ES and IA models. Given the differences in complexity and computational requirements of ES models, this effectively implies the integration of the IA model into the ES model. The technical difficulties involved in this depend on the type of IA model. For a recursive IA model, in which decisions that are simulated depend on information about current and past conditions, the IA model can be stepped forward in time along with the ES model (although technical problems might still need to be solved). However, for an intertemporal optimization IA model, in which decisions that are simulated also depend on information about future climate conditions, integration will require whole new modeling techniques (e.g. using a simplified version of an ES model (preferably the one used in the integration) to provide information of future conditions, or running the IA model in some myopic way). This type of integration allows potentially full assessment of the feedbacks between the human and earth system while retaining the full capabilities of the ES model. It could achieve the highest degree of consistency. The method, however, also involves the highest additional costs. First of all, it introduces additional complexity in the models, among others because models originating from very different scientific disciplines are coupled, each with specific terminology, focus etc. Second, as the approach typically combines only one ES model and one IA model, there is a loss of flexibility and representation of the full uncertainty space: the range of outcomes originating from using different ES or IA models is not taken into account. While theoretically many ES models have been formulated in a modular fashion with coupling software, in practice setting up the coupling is likely to be too labor intensive to allow for couplings with different models. A third risk to the approach is that the high level of complexity may also lead to intransparency.

Currently, the trend in model development seems to suggest that there is a 'natural' evolution from A towards category D-type models. However, there are clear strengths and weaknesses to each of these approaches as indicated above. In that light, it is possible to formulate a set of criteria that may be used to assess which type of integration is most suitable to address a given scientific question.

• (1)  
  If a one-way linkage between the human and the earth system is dominant because feedbacks are weak, very slow, or non-existent, category A type cooperation is the most appropriate.
• (2)  
  If interactions in both directions are known to be significant and the relevant processes can be represented by relatively simple formulations, category B is the best way forward. Most environmental change processes that take place at an aggregated scale (e.g. radiative forcing by long-lived greenhouse gases) or that can be derived from changes at an aggregated scale, fall into this category.
• (3)  
  If the main focus is on the natural system and the relevant human system processes can be represented by relatively simple formulations, category C seems the best way forward.
• (4)  
  If interactions are known to be significant (or exploration of this possibility is the goal), the relevant processes have a relatively large spatial and temporal variability and these processes cannot be adequately represented in simple models, category D is the desired option. Non-linear threshold behavior (e.g. collapse of a system under specific conditions that may become more likely with climate change) could be a reason to consider category D-type modeling as well.
• (5)  
  Uncertainty is important in the choices. If the uncertainty in involved processes and interactions is large, category A or B is probably most useful, at least as a first step allowing the flexibility to explore the full uncertainty range (certainly in light of the natural variability in ES runs). If results of these explorations suggest the possibility of a strong feedback effect, C or D analysis could be a useful next step. Exploration of uncertainty in feedbacks will be more difficult in type D calculations, given their higher computational demands.
• (6)  
  Criteria 2 and 5 considered together can also be formulated as follows: it is only useful to consider category B, C or D if potentially strong feedbacks are involved and the processes involved are rather well established.

We now discuss some IA–ES interactions that we believe warrant further exploration. This list does not intend to be complete; above all we aim to show how IA–ES cooperation can be encouraged and what forms of cooperation (A, B, C, D or a combination of them) could be most suitable given the considerations and criteria in section 2.

3.1. Impact of climate change on energy use

3.2. Impact of climate change on transport and shipping routes

3.3. Interaction between climate change and land use

3.4. Interaction between air pollution and crop growth

3.5. Interaction between climate policy and air pollution policies

3.6. Droughts, availability of water and impacts on societies

3.7. Mitigation policy responses to realized/projected climate change

3.8. Extreme and catastrophic events

3.9. Avoiding particular (regional) climate change outcomes or impacts

3.10. Other areas of research

In this letter, we have indicated that there is a need for further cooperation between the different communities studying climate change. In this context, we have specifically focused on the potential of enhanced collaboration in the form of coupling IA and ES models. As discussed in section 3, there are many examples of areas that are useful to study from both disciplines and where interactions are likely to exist. The cooperation can differ from coordinated exchange of information, the introduction of simplified representations of new processes in the models used by the IA and ES communities, and finally full model integration. These cooperation modes all have strengths and weaknesses (see table 1). The most suitable type of cooperation depends on factors like the expected strengths of the interactions, the existence of feedback mechanisms, the role of uncertainty and the advancement of science. In this light, we have discussed a set of criteria that might be used in deciding which mode of cooperation is most suitable. While full integration can potentially deal with feedbacks and consistency issues, it also leads to rather complex models and little flexibility in exploring uncertainty.

Next, we discussed a wide range of areas where IA and ES model coupling would make sense. Important areas, for instance, include the land use and land cover dynamics, the interaction between climate change and air pollution, and the role of water and droughts. Also, the impact of climate change on energy use and production, and the role of extreme events could be very relevant. Using the considerations on the most suitable forms of cooperation, we found that often simplified representations of IA and ES dynamics will be sufficient to deal with the possible linkages. In some cases, feedbacks seem to be too small to warrant more complex coupling. In other cases, linkages mostly occur via global mechanisms that can also be captured by more simple approaches. Finally, simpler approaches are more flexible in exploring uncertainty. Still, based on further cooperation between the two communities, it is likely that more complex model linkages will become more important in the future. Full integration seems to be most useful for studying local impacts, including those related to land use, as here one would benefit from the high resolution of ES models. Some of the local feedbacks are strong enough to make a difference, maybe even at higher levels of aggregation.

The research was funded by the Netherlands Organisation for Scientific Research (NWO) programme Sustainable Earth.