Large Ensemble Analytic Framework for Consequence‐Driven Discovery of Climate Change Scenarios

Abstract

An analytic scenario generation framework is developed based on the idea that the same climate outcome can result from very different socioeconomic and policy drivers. The framework builds on the Scenario Matrix Framework's abstraction of “challenges to mitigation” and “challenges to adaptation” to facilitate the flexible discovery of diverse and consequential scenarios. We combine visual and statistical techniques for interrogating a large factorial data set of 33,750 scenarios generated using the Global Change Assessment Model. We demonstrate how the analytic framework can aid in identifying which scenario assumptions are most tied to user‐specified measures for policy relevant outcomes of interest, specifically for our example high or low mitigation costs. We show that the current approach for selecting reference scenarios can miss policy relevant scenario narratives that often emerge as hybrids of optimistic and pessimistic scenario assumptions. We also show that the same scenario assumption can be associated with both high and low mitigation costs depending on the climate outcome of interest and the mitigation policy context. In the illustrative example, we show how agricultural productivity, population growth, and economic growth are most predictive of the level of mitigation costs. Formulating policy relevant scenarios of deeply and broadly uncertain futures benefits from large ensemble‐based exploration of quantitative measures of consequences. To this end, we have contributed a large database of climate change futures that can support “bottom‐up” scenario generation techniques that capture a broader array of consequences than those that emerge from limited sampling of a few reference scenarios.

Introduction

Studies of climate change impacts, mitigation, and adaptation often rely on scenarios that describe how human systems (demographics, economics, technology, etc.) will evolve in the future. These scenarios provide a backdrop against which to assess climate impacts and mitigation strategies. Shared community scenarios—scenarios that multiple users can access and use—have been important tools for inter‐study comparison and the translation of research findings across diverse disciplines. Over the last two decades the integrated assessment community has developed a series of shared scenarios including the IS92 scenarios (Leggett et al., ) and the IPCC Special Report on Emissions Scenarios (Nakicenovic, ).

In recent years, the research community has engaged in a major effort to create a new set of community socioeconomic and mitigation policy scenarios to support climate assessments (Kriegler et al., ; Moss et al., ). The resulting process created five “shared socioeconomic pathways” (SSPs), each with an accompanying set of “shared policy assumptions” (SPAs). The SSPs began with qualitative narratives about how human societies might evolve over the coming century, based on two challenge types: “challenges to adaptation” and “challenges to mitigation.” The five SSPs correspond to the four permutations of high/low challenges to mitigation and high/low challenges to adaptation with a fifth SSP that ostensibly lies in the middle. Each SSP is accompanied by an SPA that sets a mitigation policy context, including regional participation in near‐term mitigation efforts and the pricing of land‐use change (LUC) emissions. These five narrative SSP–SPA scenarios were then quantified by leading integrated assessment modeling teams, and are now available for community use and further development (Riahi et al., ). Together with the representative concentration pathways (RCPs) (van Vuuren et al., ), the SSPs and SPAs form the Scenario Matrix Framework (van Vuuren et al., ), which serves as the main technical basis for the research community to generate new global change scenarios.

Scenarios derived from the Scenario Matrix Framework are by necessity multisectoral and high‐dimensional. It is not possible, for example, to evaluate the ways temperature rise might affect society without first considering where people will live and their capacity to use air conditioning. Similarly, one cannot understand the nature of mitigation options without specifying what technologies might be available and the key drivers of emissions themselves, including population and economic growth. And for many studies needs are deeper, including demographics, education levels, energy and water infrastructures, the distribution of population, and many more important factors that influence the interaction of human societies with the Earth system.

While the SPAs and SSPs are important contributions for framing adaptation and mitigation challenges, standard implementations of the Scenario Matrix Framework (van Vuuren et al., ) are limited in how they explore the full space of relevant futures for climate change studies, considering only five socioeconomic/policy change trajectories (SSP‐SPAs) that achieve four different emissions pathways (RCPs) using perfect foresight optimization. Furthermore, the process of quantitatively abstracting the scenarios from narratives poses a challenge to our ability to understand clearly, a priori, what combinations of factors actually lead to high or low challenges. A deeper understanding of the drivers of these challenges is important for understanding the emerging shared scenarios, in modifying or customizing them for particular uses, and for designing the next iteration of shared community scenarios. This study contributes a broad exploration of the Scenario Matrix Framework and demonstrates an exploratory approach for discovering the key drivers shaping mitigation and adaption challenges.

A number of recent studies have advocated for the use exploratory modeling (Bankes, ) to generate climate scenarios (Guivarch et al., ; McJeon et al., ; Rozenberg et al., ; Weaver et al., ) rather than the more conventional narrative‐driven approach (e.g., Scenario Matrix Framework). The narrative‐driven (or scenario logics) approach begins with a priori reduction of the uncertainty space to a few aggregate scenario axes or systemic megatrends (e.g., challenges to mitigation and challenges to adaptation), which are then quantified using systems models. In contrast, exploratory modeling begins with a broad sampling of the uncertainty space then uses systems models to test the implications of different assumptions about system states and dynamics for a range of salient measures. By systematically testing across a large scenario ensemble, it may be possible to derive insights about which uncertainties are most critical to achieving or avoiding system states of interest. These insights can be used to focus future research or to inform a posteriori generated policy relevant scenario narratives (Bryant & Lempert, ; Groves & Lempert, ; Lempert et al., ). The application of exploratory modeling when studying global change is particularly appropriate because deep uncertainties (Knight, ; Walker et al., ) concerning the internal state and dynamics of the human–Earth system make a priori hypothesis falsification and assessments of scenario likelihood extremely difficult. These challenges are particularly acute for narrow explorations of the complex and high‐dimensional Scenario Matrix Framework. For an extended discussion comparing exploratory modeling and scenario logics see Greeven et al. ().

As part of their critique of the Scenario Matrix Framework, Rozenberg et al. () demonstrated exploratory modeling using an ensemble of 286 climate scenarios from the IMACLIM‐R model. That work varied sampling levels across seven sampling dimensions concerning globalization, environmental stress, carbon supply, and equity. Rozenberg et al. defined proxy metrics for the SSP mitigation and adaptation challenges as well as critical thresholds in each metric that abstract five SSP regions. They then used statistical clustering on the proxy measures to identify the key defining characteristics of scenarios in each SSP region. Assumptions concerning global income equity, labor productivity, and energy were shown to be most predictive of challenges to mitigation and adaptation, as measured by Rozenberg et al.'s metrics.

Guivarch et al. () extended Rozenberg et al.'s analysis with broader sampling of the uncertainty space (ultimately 432 scenarios) and introduced a new statistical clustering algorithm to “discover” more diverse scenario families in regions of interest. Those results showed that very different scenario families may inhabit the same SSP “challenge” region and suggests a systematic methodology for defining several scenarios for each SSP rather than a single canonical scenario. Both of these studies highlighted that clustering‐based aggregation of SSP “challenges” into a few metrics poses limits. Even a very good metric is unlikely to capture the full scope of challenges to mitigation or adaptation, and it is perhaps better to focus on specific outcomes of interest. Guivarch et al. took this approach in a second example, where they identified the shared characteristics of high emissions scenarios. More recently Marangoni et al. [] performed a multimodel sensitivity analysis of the SSP space to identify the key drivers of global emissions both with and without climate policy. Their analysis sampled five parameter types, across six Integrated Assessment Models (IAMs). Their work highlights the importance of factor interactions, wherein coinciding factor levels in different sectors can wield unexpectedly large influence compared to their impacts individually.

Building on these prior studies this study leverages the Scenario Matrix Framework to generate a large ensemble of scenarios. We then contribute a flexible analytic framework for a posteriori determination of the most important scenario dimensions that control outcomes of concern. In particular, we use the Global Change Assessment Model (GCAM) to systematically sample the SSP components, three carbon price trajectories, and four SPA mitigation policy contexts, resulting in 33,750 scenarios. As an illustrative demonstration, we show how statistical and visual analytics can then be used to identify the key scenario component drivers for regional or global metrics of interest. Beyond the specific insights derived from the illustrative scenario discovery analysis, our contributed database of GCAM runs represents a significant resource for the community to explore a broad range of adaptation or mitigation questions.

The Global Change Assessment Model

Background

The GCAM is a global integrated assessment model that couples representations of the climate, agricultural, energy, and economic systems (JGCRI, ). GCAM and its predecessors have been used in nearly every significant integrated climate‐energy assessment, including the fifth IPCC assessment report (Clarke et al., ), the development of the RCPs (Thomson et al., ), and most recently the development of the SSP‐SPA framework (Calvin et al., ). The spatial and sectoral resolution of GCAM 4.0 (available on GitHub) varies across model components. The world is divided into 32 energy/economic regions, which are themselves further divided into 283 agricultural/ecological zones. GCAM is highly modular, allowing different economic sectors (e.g., agriculture, transportation, energy) to be modeled at varying levels of detail, both globally and regionally. The regions are linked via markets for energy and agricultural commodities. GCAM is a dynamic‐recursive market equilibrium economic model, so that the prices for all modeled goods and services are adjusted across all modeled markets until supply equals demand for each 5‐year time step.

Choice between competing technologies (e.g., energy sources) is modeled in GCAM using a logit choice methodology that determines the relative market share of each technology as a function of their average relative profits and costs (Clarke & Edmonds, ). GCAM is a dynamic recursive model, so technological, fuel, and crop choices are made with only the information available in that time. However, the consequences of those decisions, such as resource depletion and capital stock, can constrain future decisions. Importantly, population, gross domestic product (GDP), and technological evolution (efficiency and cost) assumptions are exogenous to the model. The atmospheric and climatic systems are represented in GCAM using MAGICC (Meinshausen et al., ; Wigley & Raper, , , ). GCAM tracks the anthropogenic emissions of 24 greenhouse gases (GHGs), which are input into MAGICC to compute concentrations and forcing levels.

GCAM can either run with a fixed climate policy (e.g., emissions limit or carbon dioxide [CO₂] price) or solve for a least cost policy that achieves some target goal (e.g., forcing level, temperature, CO₂ concentration). Least cost solutions are derived using a Hotelling Rule approach (Hotelling, ), adjusted for ocean diffusivity (Clarke et al., ; Edmonds et al., ). Policy costs in GCAM are computed by integrating the marginal abatement cost curve, which can be understood as the total abatement cost due to loss of consumer and producer surplus (Calvin et al., ). Importantly, this metric does not capture surplus gains from avoided damages, nor does it include productivity losses (or gains) to regional GDP as a result of climate change.

Implementation of SSP–SPAs in GCAM

To implement the SSP–SPAs in GCAM, we first used the population and GDP estimates provided by IIASA (population; Samir & Lutz, ) and the Organisation for Economic Co‐operation and Development (OECD) (GDP; Dellink et al., ). For all of the other model assumptions, the qualitative storylines (O'Neill et al., ) were first reinterpreted as qualitative elements for individual sectors and technologies in GCAM (see Table ). Those qualitative elements were then translated into quantitative model parameters and assumptions, as described in Calvin et al. (). For example, agricultural productivity for each SSP was categorized as low (SSP3, SSP4 low income), medium (SSP2, SSP4 medium income), or high (SSP1, SSP5, SSP4 high income). High, medium, and low were then translated into a quantitative rate of productivity growth. In this case, medium is based on the UN Food and Agriculture Organization (FAO) projections (Bruinsma, ); high (low) is 50% above (below) the FAO estimates. These parameterizations were then combined with parameterizations for the other sectors and technologies to create an individual SSP simulation. We refer the reader to Calvin et al. () for more details and specific numbers.

Overview of SSP Implementation in GCAM (Adapted From Calvin et al. ())

Socioeconomics	Level	SSP1	SSP2	SSP3	SSP4	SSP5
	2100 Population	6.9 billion	9.0 billion	12.7 billion	9.3 billion	7.4 billion
	2100 GDP per capita	$43,306	$33,307	$12,092	$23,664	$83,496
Energy demand	Level	SSP1	SSP2	SSP3	SSP4	SSP5
	Building	Low	Med	Low	Low	High
	Transportation	Low	Med	Low	Low	High
	Industry	Low	Med	Low	Low	High
AGLU	Level	SSP1	SSP2	SSP3	SSP4	SSP5
	Food demand	High	Med	Low	Low	High
	Meat demand	Low	Med	High	Med	High
	Productivity	High	Med	Med	Med	Low
Fossil fuel costs	Level	SSP1	SSP2	SSP3	SSP4	SSP5
	Coal	Med	Med	Low	Med	Low
	Conventional gas and oil	Med	Med	Med	Low	Low
	Unconventional oil	High	Med	Med	Med	Low
Low emission energy costs	Level	Low	Med	High
	Nuclear	Prohibitive	Med	High
	Renewables	Low	Med	High
CCS	Level	Low		High
	Deployment costs	Prohibitive		Low

Each scenario element has discrete levels, corresponding to unique time series of various global and regional GCAM parameters. Some discrete levels are shared by multiple SSPs, while others have discrete levels that are unique to each SSP. For specific quantification of the various discrete levels, see Calvin et al. ()

Methodology and Exploratory Modeling

Factorial Experimental Design

This work has two objectives. First, explore the space of global futures inherent to the Scenario Matrix Framework. Second, to illustrate an analytical framework to determine the scenario conditions that are most predictive of consequential outcomes, as defined by a user's metrics of choice. The canonical SSP–SPA scenarios may miss critical combinations of events that yield unprecedented outcomes because of their focus on fixed SSP‐SPA/RCP combinations. Instead, this study employs a full factorial experimental design to contribute a broad sampling of the SSP space (Letner & Bishop, ). Although such a design may generate some unlikely scenarios, it enhances exploration and discovery of interesting scenarios that emerge from complex combinations of factors. In short, the factorial experiment shifts the expert judgments that inform scenario selection from a challenging a priori effort to a discovery process that occurs after consequential factor combinations are identified (i.e., explore first, choose later).

As illustrated in Figure , we facilitate our factorial design of experiments by aggregating the SSP assumptions into six factors, each represented by a vertical axis. The SSP factors correspond to the major elements of the SSP narratives as implemented in the various sectoral modules of GCAM by Calvin et al. (). Each factor has its own discrete treatment. Each treatment abstracts specific GCAM parameter inputs across time, regions, and assumed technologies. Figure  shows that for several of the factors, treatment levels represent a specific SSP, whereas for energy technology and CCS factors the treatment levels are shared by various SSPs. Table  provides a summary of the sampling dimensions and their associated sampling levels. A sampled “world” is generated by selecting a sampling level for each of the SSP sampling dimensions.

Fig. 1. Strategy for sampling the SSP space. The SSP assumptions are aggregated into six sampling dimensions, each having a number of discrete sampling levels composed of a distinct set of assumptions (i.e., input data and GCAM parameter values). Each sampling dimension is assigned an axis, and each sampling level is assigned a hash on an axis. A sampled SSP world as implemented in GCAM is represented by a line on the plot. If two lines intersect at an axis, those two worlds share a set off assumptions in that sampling dimension. If they do not intersect an axis at the same point, they have different assumptions in that sampling dimension. The five canonical SSPs are shown in bold color. A full factorial sampling is applied, so a world is generated for every combination of sampling level across all of the sampling dimensions, totaling 3750 sampled worlds.

In Figure , a sampled world is represented by a line intersecting all of the factors used. The five canonical SSPs are plotted as bold colored lines. When two lines intersect at an axis it means that those two worlds have a shared set of assumptions: for instance SSP2 (burnt orange) and SSP5 (brown) intersect on the Energy axis at the “Mid” level, meaning that both those scenarios have medium levels of technological improvement for renewable energy (see details in Table ). In contrast, when two lines intersect an axis at different points it means that those worlds have different sets of assumptions for that factor. For instance, all of the bold canonical SSP lines intersect the POP/GDP axis at different locations because each canonical SSP has distinct globally distributed population and GDP trajectories. Our full factorial sampling scheme samples every combination of treatment level across all of the exogenous factors used in defining the SSPs. This totals to 3750 sampled worlds, which are shown in Figure  as gray lines.

All of the sampled combinations in Figure  are self‐consistent in the sense that GCAM utilizes a unified data structure and assumptions across each of the sectoral modules. Thus, for instance, assumptions about population and productivity growth propagate consistently through the energy, land use, transportation, and other sector models. As stated before, some of the generated worlds in Figure  are unlikely. That being said, it would be extremely difficult to rigorously apply a priori subjective likelihoods to the immense combinatorial space sampled in this experiment. Moreover, it is unlikely that a consensus could be reached on the full scope of measures that should be considered or what constitutes an outcome of interest for the global climate change context. The advantage of the “bottom‐up” exploratory approach demonstrated here is that our imperfect perceptions of what is probable do not limit our exploration of what is possible. Whether some scenario is sufficiently likely to be of concern is then assessed after important system dynamics tied to consequences of interest are identified and not before (generate first and choose later).

Each of the SSP worlds illustrated in Figure  is then simulated using GCAM assuming nine different CO₂ price regimes, yielding a total of 33,750 scenarios explored in this study. Each CO₂ price regime is composed of two parts: (1) its climate policy context abstracted from the SPA space and (2) a long‐term CO₂ price trajectory. Figure  shows the sampling matrix used to construct the nine CO₂ price regimes. On the left of matrix are the narrative challenges to mitigation abstracted from the SPA space, and on the top of the matrix are the levels of mitigation effort as represented by long‐term price trajectories. The challenges to mitigation effort sampled from the SPA space have two dimensions: the level of regional fragmentation of short‐term CO₂ prices and the level of fragmentation in LUC emissions prices. Each dimension has two sampling levels (low and high). For low fragmentation in the CO₂ price regime there is a globally uniform CO₂ price starting in 2020, while at the high level there are disparate regional CO₂ prices until 2040 when a globally uniform CO₂ price is applied. Under low fragmentation in the LUC emissions price regime a small uniform LUC emissions price is applied, while at the high‐level disparate regional prices are applied where wealthy countries experience high prices and developing countries experience low prices. More details about sampled CO₂ price regimes are in the Supporting Information S1.

Fig. 2. Sampling matrix used to generate nine CO2 price regimes to be applied to each of the 3750 worlds sampled from the SSP space (see Figure ). Along the vertical axis are the four challenges to mitigation effort narratives that were abstracted from the SPA space. They are categorized by varying levels of global fragmentation in the level of near‐term CO2 prices and the level of LUC emissions prices. Along the horizontal axis are the three long‐term CO2 price trajectories, ranging from a High Price level to No Price. In the case that no CO2 price is applied (case 9), the challenge to mitigation narratives on the horizontal axis does not make sense and is not applied.

Although the computational burden of evaluating the factorial experiment illustrated in Figure  is nontrivial, it is in fact very modest compared to broader climate change projection efforts (e.g., Kay et al., ). Our experiment highlights the value to the integrated assessment community of maintaining open source and modern codes that can be effectively employed on the increasingly ubiquitous parallel high‐performance computing platforms. Moreover, our experiment highlights the need for community level data sharing initiatives that are commensurate with those of the Earth system modeling efforts for climate projections (e.g., the Coupled Model Intercomparison Project; Taylor et al., ).

Visual and Statistical Scenario Discovery

The 33,750 GCAM scenarios represent a wealth of data across many dimensions (sectoral, regional, and temporal) necessitating the use of statistical and visual scenario discovery tools. In this study, we demonstrate how to visually and statistically discover combinations of sampled SSP–SPA factors that most influence certain metrics of concern. Our approach ties to the large bodies of literature from computational scenario discovery (Bryant & Lempert, ; Kwakkel & Jaxa‐Rozen, ; Lempert et al., ), visual analytics (Andrienko et al., ; Inselberg, , , ; Inselberg & Dimsdale, ), and global sensitivity analysis (see reviews by Saltelli et al. (); Iooss and Lemaitre ()). Our proposed methodology consists of four core steps:

1. Specify thresholds or bounds on metrics that define scenario outcomes of interest.
2. Utilize parallel axes plots to visualize the subset of the scenario ensemble that meets the user‐specified outcomes of interest (“visual factor mapping”).
3. Statistically classify what factor combinations in the SSP–SPA space have a dominant impact on the outcomes of interest.
4. Use the insights from Steps 2 and 3 to aid the narration and spatial/temporal visualization of discovered consequential scenarios.

Steps 2 and 3 identify the key combinations of SSP factors (scenario elements) that are most predictive of the scenario outcomes of interest defined in Step 1. Step 4 then provides a visualization and qualitative narrative for the numerical insights derived in Steps 2 and 3. In simple cases, consequential SSP factors and key system states of interest may be identified by visual inspection alone (Step 2). However, it is often not straightforward and the use of statistical techniques may be required. In this study the Classification and Regression Tree (CART) method (Breiman et al., ) was used for scenario discovery. Alternative scenario discovery methods include the Patient Rule Induction Method (PRIM) (Guivarch et al., ; Kwakkel, ; Kwakkel & Jaxa‐Rozen, ), support vector classification (Herman et al., ), and logistic regression (Quinn, ). Lempert et al. () provide a detailed description of CART and PRIM and evaluate their performance for different test cases.

Results and Analysis

The potential costs and effectiveness of alternative mitigation strategies are important considerations in climate policy analysis. Figure  plots the radiative forcing level in 2100 versus the average global policy cost over the next century (as a fraction of global GDP) for each sampled scenario that solved (33,126 in total). The three distinct scenario clusters (colored green, red, and blue) correspond to the three long‐term CO₂ price regimes shown along the horizontal axis of Figure . As expected, for a given simulated world, higher CO₂ prices result in less warming and greater policy costs. As CO₂ prices increase, the variability of costs is greatly increased, while the variability of forcing levels is about the same (i.e., the suite of sampled GCAM runs show significant horizontal spread).

Fig. 3. 2100 forcing levels versus average global policy cost (as a fraction of global GDP) over the next century. The generated scenarios each provide unique regional time series of various socioeconomic, emission, and climate time series. As CO2 price regimes become more aggressive (from no price in green, to low prices in red, to high prices in blue), the forcing levels become lower, but also the costs increase. Two nominal 2100 climate forcing levels (denoted as CF) are highlighted with horizontal banding (± 0.1 W/m2), CF 4.5 in purple and CF 6.0 in orange. Using GCAM's target finding ability, the five canonical SSPs are paired with the two plotted CFs. The SSP–SPA 3/CF 4.5 scenario would not solve, an indication of how challenging SSP3 is, and SSP 4 stabilized below CF 6.0 in absence of any CO2 price.

The horizontal banding in Figure  corresponds to CF 4.5 (in purple) and CF 6.0 (in orange) nominal forcing levels in 2100 (± 0.1 W/m²). Strict application of the scenario matrix framework would restrict analysis to these (and two other) narrow forcing outcomes, and adjust CO₂ prices accordingly. Figure  plots nine such canonical CF/SSP–SPA combinations, derived using GCAM's target finding routine. We note that for both highlighted climate forcing levels, SSP/SPA 5 is the most costly canonical scenario (SSP–SPA 3/CF 4.5 scenario did not solve) and SSP–SPAs 1 and 4 are the least costly. This seems to confirm ex post that the SSPs were constructed correctly in the mitigation “challenge” space. By fixing the carbon price trajectories rather than target finding, our experiment results in a wider variety of forcing and cost outcomes than the standard RCP/SSP–SPA framework.

Our experiment provides a wide distribution of mitigation costs for the nominal CF 4.5 and CF 6.0 forcing levels. Figure  shows that within the CF 4.5 and CF 6.0 forcing bands there are clusters of hybrid SSP–SPA scenarios that fall outside the range of mitigation costs defined by the standard SSP–SPA scenarios. Two interesting research questions are what combinations of SSP–SPA factors result in high or low mitigation costs, and whether this depends on the forcing outcome of interest. We answer both of these questions by applying the scenario discovery techniques described in Section 3.2. In particular, scenario discovery techniques are used to identify narrative scenario families associated with high mitigation cost CF 4.5 outcomes, and both high and no mitigation cost CF 6.0 outcomes, as summarized in Table . In the following discussion outcomes of interest are user defined policy relevant conditions (e.g., high mitigation costs and CF 4.5), scenarios of interest are simulated scenarios that experience an outcome of interest, and a discovered narrative scenario family is the descriptive summary of the key defining features of the scenarios of interest revealed by the analysis (e.g., populous low productivity agriculture scenarios).

Summary of the Combined Mitigation Cost and Climate Forcing Outcomes of Interest and the Corresponding Discovered Narrative Scenario Families Highlighted in this Study

Outcome of interest	Discovered narrative scenario family
High mitigation cost, CF 4.5	Populous Low‐Productivity Agriculture Scenarios
High mitigation cost, CF 6.0	Wealthy Low‐Productivity Agriculture Scenarios
No mitigation cost, CF 6.0	High‐Inequality Low Energy Demand Scenarios
	High‐Inequality Advanced Agriculture Scenarios

High Mitigation Cost, CF 4.5 Outcomes

We start by narrowing our climate focus to CF 4.5 (purple band in Figure ), and ask which SSP factors are linked to the highest cost scenarios. Here, the 100 most expensive CF 4.5 scenarios are considered high cost and defined as scenarios of interest. Figure  applies a filter to the parallel axes plot from Figure  so that only the high cost CF 4.5 scenarios of interest are retained. It is particularly noteworthy if all of the retained scenarios either pass through or avoid some point in the experimental design, because that indicates that all scenarios of interest share that SSP feature or that that feature precludes a scenario from the scenarios of interest. Two key features emerge from the visual factor map in Figure . All high cost CF 4.5 scenarios share SSP3 population and GDP assumptions, and nearly all (96%) share SSP3 agriculture and land use (AGLU) assumptions. The four high‐cost scenarios which do not share SSP3 AGLU assumptions share SSP2 assumptions. All of our visual factor mapping results are confirmed using the CART algorithm described in Section 3.2. We will refer to the scenarios of interest highlighted in Figure  as the Populous Low Productivity Agriculture Scenarios.

Fig. 4. Populous Low‐Productivity Agriculture Scenarios. The full experimental design in Figure  has been filtered so that only the 100 highest cost CF 4.5 scenarios are retained. All of the canonical scenarios have been filtered, but are retained in transparency for reference. All of the retained scenarios pass through the Pop/GDP axis at the SSP3 level, and 96% pass through the AGLU axis at the SSP3 level. This indicates that the vast majority of high cost CF 4.5 scenarios share SSP3 population, GDP, and AGLU assumptions.

As shown in Figure , the Populous Low Productivity Agriculture Scenarios share SSP3 population and GDP assumptions. SSP3 population assumptions are characterized by a significant growth in global population, reaching 12.7 billion in the year 2100. As shown by Figure a, this growth is mainly concentrated in developing countries, and is particularly acute in sub‐Saharan Africa and parts of the Middle East (Samir & Lutz, ). Population growth in the developing world is accompanied by a decline in population in the OECD countries due to low fertility and low rates of migration (Samir & Lutz, ). Furthermore, the Populous Low Productivity Agriculture Scenarios see stalled convergence in per‐capita income between developed and developing countries, resulting in an international distribution of wealth in 2100 that is similar to present levels (Dellink et al., ).

Fig. 5. SSP3 Pop/GDP and AGLU assumptions. All of the Populous Low‐Productivity Agriculture Scenarios share SSP3 population and GDP assumptions, characterized by a large growth in the global population. (a) The regional distribution of the population growth in percent change relative to 2015. While much of Africa and the Middle East see significant population growth, some parts of the OECD see population decline. Most of the Populous Low‐Productivity Agriculture Scenarios share SSP3 AGLU assumptions, which is characterized by limited improvements in agricultural yields. (b) The regional relative improvement in maize yields compared to the reference SSP2 improvement trajectory. While maize yields improvements are poor globally, yields in Sub‐Saharan Africa are particularly poor. Relative improvements for other crops are similar. The combination of high population growth (a) and poor agricultural yields (b) results in significant LUC emissions.

Figure  also shows that most (96%) of the Populous Low Productivity Agriculture Scenarios share SSP3 AGLU assumptions. SSP3 AGLU assumptions are characterized by limited improvements in agricultural productivity globally, but particularly in southern Africa and northern South America (Popp et al., ). Figure b shows the geographic distribution of improvements in maize yields in 2100 under SSP3, relative to the FAO reference scenario. Similar trends exist in other crops. In the Populous Low Productivity Agriculture Scenarios, the limited yield improvement is driven by limited development of new agricultural technologies, and limited transfer of technologies to developing countries (Popp et al., ). It is important to note that the SSP narratives are independent of climate outcomes, so Figure b does not reflect the potential impact of climate change on crop productivity.

The result of these combined factors (significant population growth, low wealth, low agricultural productivity), is that a lot of new land must be brought into production to feed the growing population, which results in massive LUC emissions (Fujimori et al., ). To curb warming in the Populous Low Productivity Agriculture Scenarios highlighted in Figure , those emissions must be offset through more expensive technological means (i.e., more efficient consumption, cleaner generation). The combination of high population and low agricultural productivity was identified by Fujimori et al. () as one of the key challenges in SSP3. Furthermore, agricultural intensification has been shown to be one of the most cost‐effective means of climate mitigation (Burney et al., ), and improvements in poorer countries are expected to be particularly important to avert significant emissions over the next 35 years (Tilman et al., ). Our findings suggest that agriculture research and the deployment of advanced farming techniques are critical in averting expensive mitigation scenarios. The findings in Figure  show that high mitigation costs at the CF 4.5 level are characterized by high population, low wealth, and limited agricultural productivity. However, it is unclear whether the same SSP factors will always be tied to high mitigation challenges regardless of climate outcome, as the Scenario Matrix Framework implicitly assumes.

High Mitigation Cost, CF 6.0 Outcomes

We next shift focus to scenarios that yield a forcing near CF 6.0 in 2100, highlighted in the orange band in Figure . Figure  shows that a range of CO₂ prices yield scenarios that achieve CF 6.0 forcing levels. Our next scenario analysis focuses on CF 6.0 scenarios that emerge under the high long‐term CO₂ price regime (blue scenarios falling in the orange band in Figure ). These 163 scenarios represent the highest cost CF 6.0 scenarios in our ensemble (costs >0.4% GDP). They are also scenarios in which high CO₂ prices result in relatively limited climate mitigation. In Figure  a filter is applied to the parallel axes plot from Figure  so that only the high CO₂ price CF 6.0 scenarios are retained. Again, the pattern is very easy to identify visually: all of the retained scenarios share SSP5 population and GDP assumptions, and SSP3 AGLU assumptions. We will refer to these scenarios as the Wealthy Low Productivity Agriculture Scenarios.

Fig. 6. Wealthy Low‐Productivity Agriculture Scenarios. The full experimental design in Figure  has been filtered so that only the high long‐term CO2 price CF 6.0 scenarios are retained. All of the canonical scenarios have been filtered, but are retained in transparency for reference. All of the retained scenarios pass through the Pop/GDP axis at the SSP5 level and pass through the AGLU axis at the SSP3 level. This indicates that the high‐price CF 6.0 scenarios share SSP5 population and GDP, and SSP3 AGLU assumptions. Importantly, at the CF 6.0 level the most extreme sampled scenarios arise from hybrids of the SSPs rather than from a single canonical SSP.

All of the Wealthy Low Productivity Agriculture Scenarios share SSP5 population and GDP assumptions. SSP5 is characterized by rapid economic development driven by fossil fuel consumption, leading to declining poverty and decreased birth rates in developing countries that make climate adaption more possible (Kriegler et al., ). In fact the population and GDP dynamics of SSP5 are important defining features of SSP5. Under SSP5 population assumptions, the global population peaks around 8.6 billion around 2050 and declines to about 7.4 billion in 2100 (Samir & Lutz, ). Figure a shows the distribution of the global population in 2100 relative to 2015 levels in the Wealthy Low Productivity Agriculture Scenarios highlighted in Figure . While population growth is slowed in Africa, parts of the OECD see significant growth, due to increased fertility and migration (Dellink et al., ; Samir & Lutz, ).

Fig. 7. SSP5 population and GDP assumptions. All of the Wealthy Low Productivity Agriculture Scenarios share SSP5 population and GDP assumptions, which are characterized by a global population that peaks at 8.6 billion in 2050, and reduction in global poverty. (a) The regional distribution of the population growth in 2100 as a percent change from 2015 levels. Population growth in the developing world is modest or negative, while parts of the OECD greater levels of growth due, in part, to migration. (b) The regional increase in per capita GDP (PPP) relative to 2015 levels. While each region sees an increase in PPP, this is most pronounced in much of Sub‐Saharan Africa (80‐fold increase in East Africa). The combination of poor agricultural yields, modest population growth, and higher levels of consumption that accompany wealth (as assumed by GCAM) results in expensive CF 6.0 scenarios, and renders the High CO2 prices ineffective.

SSP5 GDP assumptions include a significant growth in global GDP, with significant improvements in per capita GDP in the developing world (Dellink et al., ). Figure b plots the relative improvement in per capita GDP in 2100 compared to that in 2015 levels for the Wealthy Low Productivity Agriculture Scenarios highlighted in Figure . While per capita incomes increase significantly globally, Eastern Africa sees an 8000% increase. This results in a significant increase in food demand across much of the developing world, particularly for animal products (Bodirsky et al., ; Kriegler et al., ). In the canonical SSP5 scenario, the impact of this increased food demand on LUC is somewhat ameliorated by widespread dissemination of advanced agricultural technologies to improve yields. However, this is not the case for the high cost CF 6.0 scenarios highlighted in Figure  that all share pessimistic SSP3 AGLU assumptions (see Figure b). Here, large LUCs occur, not because of massive population growth (as in previous example), but because of changes in income lead to changes in dietary preference. In fact, these conditions account for the entire upper limb of the high price cloud of scenarios in Figure  (i.e., high CO₂ price scenarios having 2100 forcing in excess 6 W/m²). These results are notable, in part because they show that the extreme scenarios of interest (expensive CF 6.0 and ineffective high CO₂ price scenarios) emerge as hybrids rather than a single SSP–SPA scenario.

No Mitigation Cost, CF 6.0 Outcomes

Figure  highlights the Wealthy Low Productivity Agriculture Scenarios, which achieve CF 6.0 forcing levels despite a high CO₂ price. Figure  shows that many scenarios also achieve CF 6.0 nominal forcing in 2100 in absence of any CO₂ price (i.e., the green scenarios falling in the orange CF 6.0 band). Our final scenario analysis focuses on these no cost CF 6.0 scenarios. Figure  applies a filter to the parallel axes plot from Figure  so that only the no‐cost scenarios of interest are retained. It is immediately apparent that the controlling SSP interactions are a bit more complex than in Figures  and , but we can make two easy observations. First, we note that it is not possible for a scenario with no CO₂ price to achieve CF 6.0 forcing levels with SSP3 fossil fuel extraction costs. SSP3 assumes that fossil fuel extraction costs (especially coal) will fall dramatically over the next century (see Figure 11b) (Calvin et al., ). Second, we note that all no cost CF 6.0 scenarios share either SSP3 or SSP4 demographics, which are similar except that SSP3 has more extensive population growth and SSP4 has more wealth concentration in the OECD (Dellink et al., ; Samir & Lutz, ).

Fig. 8. No cost CF 6.0 scenarios. The full experimental design in Figure  has been filtered so that only the no‐cost CF 6.0 scenarios are retained. All of the canonical scenarios have been filtered, but are retained in transparency for reference. All of the retained scenarios pass through the Pop/GDP axis at either the SSP3 or SSP4 levels, and avoid passing through the fossil fuel extraction axis at the SSP3 level. This indicates that all of the no‐cost CF 6.0 scenarios avoid SSP3 fossil fuel extraction costs, and they all share either SSP3 or SSP4 population and GDP assumptions. CART is used to identify two families of retained worlds: the High‐Inequality Low Energy Demand scenarios in pink and the High‐Inequality Advanced Agriculture scenarios in light green. These families are characterized by either the SSP1 energy demand assumptions or the SSP1 population and GDP assumptions.

While a rapid increase in population and sustained poverty in Africa was associated with high mitigation cost CF 4.5 outcomes (in which a large CO₂ price was applied), it is also associated with CF 6.0 that have no mitigation cost. In the absence of a CO₂ price, the best climate outcomes result from entrenched poverty in Africa because it suppresses energy consumption from carbon emitting technologies. If a price is applied, the opposite is true: expensive outcomes result from continued poverty and large population growth in Africa. This highlights the inherent danger of a priori specifying some assumption as high or low challenge in absence of a long‐term climate policy context or a designated stabilization level, both of which are exogenous to the SSP‐SPA “challenge” framework. However, entrenched poverty in the developing world will obviously present many other global challenges and risks.

We are able to draw deeper insights about the no cost CF 6.0 scenarios by using CART to classify the scenarios highlighted in Figure . CART results in a classification tree that can distinguish between distinct “families” of similar scenarios. Figure  is a classification tree representation of the CART results. For clarity, nodes representing leaves that contain no scenarios of interest are plotted in gray, and the tree has been truncated.

Fig. 9. Truncated CART classification tree for no cost CF 6.0 scenarios. The CART algorithm is applied to the input factor space to identify distinct narrative scenario families of no‐cost CF 6.0 scenarios. Each terminal node in the tree, called a leaf, represents a scenario family. Leafs for families with no no‐cost CF 6.0 scenarios are plotted in gray. The progress of the CART algorithm is tracked from left to right. Each split represents a partitioning of the data. The dimension of partition is labeled in black with a vertical hash. Each node lists each corresponding family's most recent membership condition and coverage (c) (the fraction of no‐cost CF 6.0 scenarios on that node). The first partition is in the energy demand dimension, and reveals that 80% of the no‐cost CF 6.0 scenarios share SSP1 energy demand assumptions, hereafter referred to as the High‐Inequality Low Energy Demand Scenarios and labeled in pink. The second partition reveals that the remaining 20% of scenarios that do not have SSP1 energy demand assumptions (!SSP1) all share SSP1 AGLU assumptions, hereafter referred to as High Inequality Advanced Agriculture Scenarios and labeled in green.

The first CART partition illustrated in Figure  is applied to the energy demand factor dimension, and reveals that 80% of the no cost CF 6.0 scenarios share SSP1 energy demand assumptions. We will refer to these as the High Inequality Low Energy Demand Scenarios, plotted in pink. The second CART partition, on the bottom branch, reveals that the remaining 20% of the scenarios of interest all share SSP1 AGLU assumptions. We will refer to these as the High Inequality Advanced Agriculture Scenarios, plotted in light green. Thus CART has revealed two families of no cost CF 6.0 worlds. The most optimistic CF 6.0 scenarios emerge as SSP hybrids rather than from a single “low challenge” SSP‐SPA narrative.

The High Inequality Low Energy Demand Scenarios all share SSP1 industry, transportation, and building efficiency assumptions, which dictate that energy demand in each sector is low. This is due to two factors: an increase in efficiencies and a change in societal consumptive behavior (van Vuuren et al., ). In the transportation sector specifically the reduced demand is due to the wider adoption of public transit, adoption of higher efficiency modes of transit, and reduced willingness to travel over time (Bauer et al., ; van Vuuren et al., ). Figure  shows the global distribution of energy consumption reduction in the transportation, building services, and industrial sectors under SSP1 assumptions for those sectors relative to reference SSP2 levels. While consumption reductions are almost always global, the most dramatic reductions in all sectors are in Africa. These efficiencies and behavioral changes, when paired with low consumption in Africa, and relatively expensive fossil fuels (i.e., not SSP3 fossil fuel extraction costs) result in relative positive climate outcomes in absence of a carbon price. These results are similar to the finding by Marangoni et al. () that future CO₂ emissions under the SSP framework are most sensitive to assumptions about GDP and energy demand.

Fig. 10. SSP1 energy use reductions in the Building (a), Transportation (b), and Industrial (c) end‐use sectors compared to reference scenario (SSP2). SSP 1 assumes a reduction in energy demands from increased efficiencies and changes in consumptive patterns. To isolate the effect of the SSP1 energy demand assumptions, the aggregate energy use in each sector from a reference scenario (SSP2) with no CO2 price is compared to an identical scenario, but with SSP1 Energy Demand assumptions. Energy use reductions are reported as percent decrease from reference levels. While reductions are modest in most of the developed world, reductions in in energy use are most pronounced in the developing world, most notably in Sub‐Saharan Africa.

The High Inequality Advanced Agriculture Scenarios all share SSP1 AGLU assumptions, and either SSP1 or SSP4 fossil fuel extraction costs. These scenarios see vast improvements in agricultural yields globally, but particularly focused in sub‐Saharan Africa. As an example, Figure a plots the global distribution of Maize yields at the end of the century, relative to the FAO reference scenario. While improvements in excess of 80% are focused in sub‐Saharan Africa and northern South America, parts of North America, Europe, and China see improvements in excess of 30%. Other crops have similar trends. Agriculture, land use, and LUC account for nearly 25% of GHG emissions, in large part thanks to deforestation (Tubiello et al., ). SSP1 AGLU assumptions see rapid development and deployment of yield‐improving technologies across the developing world, thus averting much of the land use emissions that might be expected to feed growing populations. In addition to improved agriculture, these scenarios see relatively expensive fossil fuel extraction costs (i.e., SSP1 or SSP4 levels). Figure b plots the extraction cost of coal for each SSP through the year 2100. Under SSP1 and SSP4 levels, coal extraction costs either remain unchanged or decline slightly over the next century, meaning that coal is less competitive with renewable energy sources. Similar trends exist for petroleum and gas production.

Fig. 11. Key characteristics of High‐Inequality Advanced Agriculture worlds. High‐Inequality Advanced Agricultural Scenarios see global improvement in crop yields due to the development and dissemination of advanced technologies and efficient practices, and relatively expensive fossil fuels. (a) The geographic distribution of maize yield improvements over the next century in SSP1. Generally, yields improve globally, but the most substantial improvements are in Sub‐Saharan Africa. Other crops show similar trends. (b) The extraction cost for coal through the year 2100. All High‐Inequality Advanced Agricultural Scenarios share either SSP1 or SSP4 fossil fuel assumptions, which see little or no decrease in the cost of coal extraction over the century. Trends are similar for other fuels.

While the findings about individual factors' importance are well supported in the literature, the important contributions of the exploratory modeling approach are in revealing unexpected combinations of factors that return good or bad results, and in highlighting the most critical SSP–SPA factors that define scenarios of interest. The hybrid scenario families in Figures  and represent consequential combinations of optimistic and pessimistic scenario assumptions that are well beyond the scope of the standard SSP scenarios, and would be difficult to elicit a priori. It seems obvious that agriculture is important to climate abatement, yet this is not reflected in much of the research (Burney et al., ). Our findings highlight the relative importance of agricultural productivity compared to other factors like energy technology and CCS.

Conclusions

The objective of scenario generation is to inform the decision‐making process by exposing competing action strategies to various states of the world to evaluate their efficacy and robustness. Scenarios must be sufficiently diverse, extreme, and detailed to ensure that important failure mechanisms are captured. This can be challenging for complex integrated systems, like the human–Earth system, because it is not clear which combinations of factors are important, how successes or failures should be measured, or how those questions depend on the climate and decision context. Even more fundamentally, in studies of global change it is unclear how “feasibility” should be defined or bounded.

To address these challenges, this work proposes an exploratory modeling extension of the new Scenario Matrix Framework to generate decision‐relevant scenarios. The SSP and SPA uncertainty space is broadly and thoroughly sampled to generate a large ensemble of scenarios. Scenario discovery tools, including advanced data visualization and statistical categorization, are used to identify the scenario elements most tied to several outcomes of interest. This represents an a posteriori approach, in which decision relevant scenarios are identified (or discovered), as opposed to the more traditional a priori approach in which scenarios are specified to achieve an effect (e.g., high or low challenges).

This work suggests that relying solely on a few, expert elicited, a priori specified scenarios is ill advised. We show that the most extreme and decision relevant scenarios often arise as hybrids of the various SSPs rather than from a single deep narrative SSP. For instance, the highest cost CF 6.0 outcomes arise from hybrid scenarios combining growing wealth in Africa (SSP5) and lackluster agricultural productivity (SSP3), a combination of factors not considered by the Scenario Matrix Framework. We also show that the same scenario characteristics can be associated with both good or bad outcomes, depending on the climate policy context and the forcing outcomes of interest. Specifically, we show that entrenched poverty in Africa has a mitigating effect on warming in the absence of a CO₂ price, but also greatly increases mitigation costs if a CO₂ price is applied. Based on these findings we conclude that direct use of the SSP, SPA, and CFs as initially proposed is likely insufficient for decision‐making. Instead, we have demonstrated an alternative approach, in which scenarios could be tailored to a specific decision context, based on multiple, time‐varying metrics.

The scenario matrix framework represents a significant and sustained intellectual effort by the scientific community. In our opinion, this framework should be viewed as a starting point rather than a terminus. By using the scenario matrix framework as its foundation, the database generated by this experiment could serve as the start of an extensive multimodel database of scenarios as envisioned by Rozenberg et al. () in the inaugural SSP special issue. Such a database will serve as a tool to empower decision makers to select the most relevant scenarios for their decision context.

Acknowledgments

Portions of this work were supported by the National Science Foundation through the Network for Sustainable Climate Risk Management (SCRiM) under NSF cooperative agreement GEO‐1240507. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the US National Science Foundation. The authors are not aware of any real or perceived conflicts of interest. Data and visualization code used in this analysis are available on Github (https://github.com/JRLamontagne/Factorial_SSP‐SPA_Exploration) in the Master branch.