Scenario Discovery Analysis of Drivers of Solar and Wind Energy Transitions Through 2050

Deep human‐Earth system uncertainties and strong multi‐sector dynamics make it difficult to anticipate which conditions are most likely to lead to higher or lower adoption of renewable energy, and models project a broad range of future solar and wind energy shares across future scenarios. To elucidate these dynamics, we explore a large data set of scenarios simulated from the Global Change Analysis Model (GCAM), and use scenario discovery to identify the most significant factors affecting solar and wind adoption by mid‐century. We generated a data set of over 4,000 scenarios from GCAM by varying 12 different socioeconomic factors at high and low levels, including assumptions about future energy demand, resource costs, and fossil fuel emissions paths, as well as specific technology assumptions including wind and solar backup requirements and storage costs. Using scenario discovery, we assess the most important factors globally and regionally in creating high fractions of solar and wind energy and explore interconnected effects on other systems including water and non‐CO2 emissions. Globally and regionally, we found that solar and wind‐related technology costs were the primary drivers of high wind and solar energy adoption, though a few regions depend heavily on other parameters like carbon capture and storage costs, population and gross domestic product trajectories, and fossil fuel costs. We also identify four key paths to high solar and wind energy by mid‐century and discuss their tradeoffs in terms of other outcomes.

10.1029/2022EF003442 2 of 12 potential to impact others (Graham et al., 2020), and policy interventions like renewable portfolio standards are subject to political and societal forces that are difficult to predict or reliably characterize probabilistically.
Previous studies have found a variety of different controls impact the adoption of these energy sources. These factors include renewable portfolio standards, renewable energy share, energy intensity, economic activity, carbon taxes, technology and energy costs, as well as individual household income (Bird et al., 2005;Jaxa-Rozen & Trutnevyte, 2021;Sadorsky, 2021).
By mid-century, across a range of future scenarios commonly used in climate analysis, estimates of the wind and solar share of electricity generation vary from less than 1% to nearly 40%. By the end of the century that estimated range widens to 1%-75% (Riahi et al., 2017). Higher shares of solar and wind energy are associated with a variety of outcomes across multiple sectors beyond CO 2 emissions savings. Increased renewable energy is associated with a reduction in air pollution and thus improved public health outcomes, as well as lower water consumption (Luderer et al., 2019;Lukanov & Krieger, 2019;Qiu et al., 2022;Wiser et al., 2016). Switching to renewable energy also supports local and regional energy independence and reduces exposure to fuel price volatility, compared to fossil energy sources (Morthorst & Awerbuch, 2009). Fossil fuel savings and job creation have also been attributed to renewable energy adoption (Ortega-Izquierdo & Río, 2020). These multisectoral outcomes can then have effects throughout the interconnected human-Earth system.
While there have been some efforts to probabilistically quantify uncertainty about future energy systems (e.g., Morris et al., 2022), the multisectoral complexity of renewable energy drivers also lends itself to a deep uncertainty representation. Deep uncertainty refers to outcomes where the appropriate system models to represent the relevant interactions, the probability distributions around key variables, and the valuation of the desirability of alternative outcomes are all either unknown or not agreed upon (Lempert, 2003).
A powerful approach to analyze deeply uncertain factors is scenario discovery. This technique allows us to explore a wide parameter space and identify individual and nonlinear combinations of factors driving outcomes of interest, in our case, high renewable adoption (Lamontagne et al., 2018;Moallemi et al., 2021). Applying scenario discovery methods to models that include multiple sectors and their interactions can begin to address these complex questions. Previous analyses have used scenario discovery techniques in a variety of applications, such as testing the robustness of outcomes to uncertain factors, exploring the impacts of land and water scarcity, and understanding energy futures (Dolan et al., 2021(Dolan et al., , 2022Halim et al., 2016;Moksnes et al., 2019).
This study builds a large ensemble to identify key drivers of solar and wind adoption by mid-century using the multi sector Global Change Analysis Model (GCAM). GCAM is a human-Earth system model, including explicit representations of energy, agriculture, land use, and water, as well as atmosphere and climate systems through coupling with the reduced complexity climate model Hector. GCAM also includes detailed energy and technology sectors, allowing for the exploration of various factors relevant for solar and wind energy adoption. It is a global model with discrete regions whose sectors and interactions are explicitly modeled (see Model Description for more detail). In this analysis we vary key parameters in GCAM and map them back to the resulting projected levels of wind and solar electricity deployment across our scenarios.

Scenario Development
To explore the adoption of solar and wind energy in GCAM, we constructed an ensemble of model scenarios specifically varying parameters in GCAM shown to be (directly or indirectly) relevant to rates of adoption in past work with GCAM, including emissions constraint scenarios known to encourage renewable adoption in GCAM (Bauer et al., 2017;Calvin et al., 2019;Iyer et al., 2015;Kanyako & Baker, 2021;Vinca et al., 2018). This results in an ensemble that features higher than present day wind and solar energy production in every case ( Figure S3 in Supporting Information S1). While this ensemble of scenarios is not exhaustive, it is intended to be sufficient for uncovering the tradeoffs across sectoral metrics in different regions under high global solar and wind energy adoption.
While numerous factors affect solar and wind energy adoption, we focus in this analysis on a set of 12 parameters selected based on expert opinion of the most relevant inputs available to adjust in GCAM (Table 1). These are energy costs, including fossil fuel and carbon storage resource costs; nuclear, wind, and solar capital overnight costs, including both technology and storage costs for wind and solar energy; backup requirements for intermittent wind and solar technology; a combination of changes in population and gross domestic product (GDP), which create overall changes in energy demand; electrification of building, industry and transportation sectors; as well as limits to bioenergy use, and CO 2 emissions constraints. By varying each at high and low levels, this results in a 4,096 scenario ensemble. Of these scenarios, GCAM was unable to find a numerical solution to 107, so our final results use a set of 3,989 scenarios.
For 8 of the 12 parameters, we used high and low values from the standard GCAM distribution, previously developed to support representation of the Shared Socioeconomic Pathways (http://jgcri.github.io/gcam-doc/ssp.html; Bauer et al., 2017;Calvin et al., 2017). For the remaining parameters, we created high and low scenarios from expert input, which are described in more detail below.
We selected two emissions constraints from the IIASA database (Huppmann et al., 2019) based on high and low overshoot of a two-degree Celsius end-of-century target ( Figure S1 in Supporting Information S1). In addition to these criteria, we selected two pathways with limited negative emissions, since our ensemble also involved constraining bioenergy in half of the runs, and we do not include direct air capture so the model relies on bioenergy as the primary technology to achieve negative emissions. By mid-century these emissions pathways differ by 2.4 Pg C/year but their associated temperature trajectories, as estimated by GCAM's coupled reduced complexity climate model Hector (Hartin et al., 2015), converge to within 0.06°C by 2100 ( Figure S1 in Supporting Information S1). Across the full set of scenarios in our ensemble, variation in unconstrained land use change and non-CO 2 emissions due to changes in the other ensemble parameters created a wider spread in end-of-century temperatures, of nearly 1.9-2.1°C ( Figure S2 in Supporting Information S1). These constraints act on GCAM's carbon dioxide emissions from energy and industry and are enforced through a price on carbon dioxide emissions. High and low labels for these emissions constraints are assigned based on the emissions values in 2050, the year we focus on for our analysis. In order to meet more strenuous constraints, more low and no carbon energy are needed in the energy portfolio, making more wind and solar more appealing. Note. Exact implementation details can be found in the corresponding input files, available in the accompanying data repository.

Table 1 Description of Each Parameter Sampled to Create the 4,096-Scenario Ensemble
Our bioenergy constraint was set to 100 EJ, a value chosen based on expert opinion. This constraint taxes additional second generation bioenergy in GCAM, discouraging the model from increasing bioenergy above the specified constraint. First generation bioenergy components, such as biodiesel or corn for ethanol, are not limited through this constraint. The "high" scenario for bioenergy is an unconstrained model run. By constraining bioenergy, we limit the model's options for meeting the given carbon emissions constraints, which may drive up solar and wind energy adoption.
High and low backup scenarios were generated by setting the backup capacity requirements in GCAM, which are determined as a function of the share of intermittent wind and solar of all electricity generation. The "high" ("low") scenario is driven by high (low) backup requirements, so that higher (lower) backup capacity is required by the model at lower (higher) shares of wind and solar electricity generation. Higher backup requirements make investing in intermittent renewable energy more costly.
Finally, we created a high electrification scenario by increasing the model's preference for electricity as a fuel in several subsectors within the building, industry, and transportation sectors in GCAM. The "low" scenario for this parameter corresponds to no modification of electrification from a base model run. Higher levels of electrification allow more flexibility for wind and solar energy to play a role in the energy generation in a given scenario, as these power sources are more easily substituted for electricity generation, than in other energy sectors.

CART Analysis
Classification and Regression Trees (CART) is a common tool in large-ensemble scenario analyses (Bryant & Lempert, 2010;Dolan et al., 2021;Lamontagne et al., 2018) and can be applied in either a regression or classification mode. In regression, CART identifies those values of the input parameters that are most predictive of a continuous response, such as solar and wind penetration (see, e.g., Dolan et al., 2021). In classification mode, CART identifies those conditions that are most predictive of particular discrete outcomes (e.g., solar and wind penetration above 80%) represented by binary indicator variables (see, e.g., Lamontagne et al., 2018).
We utilize both modes of CART analysis in this study to answer different research questions. In the case of classification, CART works by sequentially making binary splits of the data, defined by ranges of the input variables, in order to maximize node purity (the homogeneity of each node with respect to its corresponding category) in the case of the two groups resulting from a split. Similarly, in the case of regression, CART works by sequentially dividing the data through binary splits to minimize the variance in the resulting groups while maximizing variance between them.
In this study, we utilize the R package R-part (version 4.1-15, https://cran.r-project.org/web/packages/rpart/ index.html) implementation of CART. Using CART regression, we identify the leading drivers associated with the level of wind and solar energy production in 2050 globally and in different regions. This tells us the effect different GCAM input parameters have on overall wind and solar adoption across the entire range of wind and solar energy levels seen across all of our scenarios.
In contrast, we applied CART classification to identify the specific combinations of GCAM parameter values associated with the highest levels of wind and solar adoption, a related but distinct analysis speaking to a particular range of the outcome space rather than the full range as in regression. We specifically looked at the top 5% of global wind and solar energy penetration in our classification analysis, which is a somewhat arbitrary selection, but as the data lacks a high natural break, 5% is an approximate point to delineate the top wind and solar energy scenarios (see the Discussion for more on this choice of "top" set of scenarios). To address overfitting, CART trees are pruned based on the cross-validation error.

Model Description
For this analysis we use GCAM, an open-source, freely available, multisector model developed and maintained by the Pacific Northwest National Laboratory. It is a dynamic recursive partial equilibrium model that numerically simulates a wide range of markets worldwide, representing key interactions between climate, energy, land, water, and economic systems. Model scenarios are driven by external assumptions including socioeconomic drivers, technology costs and efficiencies, as well as government policy . GCAM divides the world into 32 socioeconomic regions, with further resolution present in certain sectors, such as water (235 hydrological basins) and agriculture (384 land units formed by intersecting basins with socioeconomic regions).
Our ensemble was specifically run on a modified version of GCAMv5.4 (Bond-Lamberty et al., 2021) with a few specific upgrades (see the Open Research section for accessing a copy of this model version). The additional updates included in the model version we use are an updated residential floor space demand function, improved handling of water market solutions, and a handful of technical improvements that do not affect model results.
These updates are now included in the GCAMv6.0 release and are described in more detail in the model documentation (https://jgcri.github.io/gcam-doc/).

Global and Regional Drivers
Wind and solar fractions of electricity generation increase substantially by mid-century in our ensemble, from around 10% in 2020, to 35%-63% globally across all scenarios ( Figure S3 in Supporting Information S1). In the top 5% of scenarios (estimated based on the total combined wind and solar fraction of electricity generation in 2050) this range narrows to 58%-63% solar and wind electricity generation, with strong regional heterogeneity (Figure 1a). By 2050 in this top set of scenarios, much of south and southeast Asia, Africa, and Oceania get more than 60% of their electricity from wind and solar, while Russia and some former Soviet countries are estimated to get close to 30%. With some of the largest renewable fractions, Central America, Pakistan, South Africa, and eastern Africa have a higher than 80% share of wind and solar electricity generation by mid-century. Over the 30 year period from 2020 to 2050, solar electricity generation increases strongly in the top 5% of scenarios in China and India, while the US and EU see stronger growth in wind (Figure 1b). Generally, solar electricity generation predominates in the southern hemisphere, except in Argentina, while wind electricity plays a stronger role in the electricity sector in regions in the northern hemisphere. Four regions (United States, EU-15, China, and India) make up nearly 36% of the total global wind and solar electricity generation in 2050. At the global scale, based on CART variable importance scores, the top drivers of high combined wind and solar energy fractions in 2050 are population and GDP, wind technology costs, and solar storage costs (Table S1 in Supporting Information S1). The next highest drivers are fossil fuel costs, bioenergy constraints, and carbon capture and storage costs, while solar technology costs were much less influential and wind storage costs played an almost negligible role. At the regional level, looking across all scenarios, wind and solar technology costs, solar storage costs and carbon capture and storage (CCS) costs stand out as the most common strongest determinants of combined solar and wind electricity fractions; however, there is wide variation among regions (Figure 2). Regions like Japan, Canada, and Taiwan are primarily driven by solar storage costs, while much of Africa is driven by solar technology costs. Several countries, including Central America and Southeast Asia, see carbon capture and storage costs playing a larger role in controlling renewable energy. Population and GDP, which affect overall energy demand, show up as the strongest driver for the European Free Trade Association, and a somewhat influential driver in several other areas. South Asia and Australia also see an impact from fossil fuel costs, which otherwise plays a minor role in regional profiles. Across all regions, however, backup electricity generation requirements, nuclear costs, emissions constraints, electrification, and wind storage costs play a relatively minor role.

Different Paths to the Top
Using CART as a classification algorithm on the global wind and solar fraction of electricity generation, we identify combinations of drivers that lead to scenarios being in the top 5% (200 scenarios from our successfully solved 3,988). Nearly 90% of scenarios in this top category had low wind and solar technology costs and high carbon capture and storage costs (176 out of 200), and more than 50% included a bioenergy constraint (Figure 3). Within the other 10% or so of top-adoption scenarios a few reached high wind and solar shares when wind technology costs were high but were compensated by lower solar technology costs or vice versa, and a few others had high wind and solar energy when CCS was cheap as long as wind technology costs were also low.
All other scenarios in this top set can be divided into four key paths of parameter combinations illustrated in Figure 3. In addition to low wind and solar technology costs and high CCS costs, path #1 is defined by a low population and GDP scenario, in contrast to the remaining paths which all have high population and GDP. This is a less likely path toward high renewables, with only 28 of the 256 scenarios with this combination of parameters falling into the top 5%. Of these 28, 24 had a common set of additional parameters that combined to create high renewable energy fractions: high fossil fuel costs, low solar storage costs, and constrained bioenergy. As the unique factor in this path is low population and GDP, this is used as a shorthand for the path throughout the following text.
Path #1 makes up the smallest fraction (12%) of the top 5% of scenarios and 8 of the 32 scenarios with this combination of parameters ended up with lower wind and solar penetration. Lower population and GDP reduces overall pressure on the energy system and thus fossil fuel emissions, reducing the need to convert to wind and solar energy as quickly to meet either emissions target. This reduced pressure to convert to renewables means that when any of the factors on this path are not true (fossil fuels are less expensive, bioenergy is more available, or solar storage is more expensive) then the model ends up at more intermediate levels of solar and wind energy.
Under high population and GDP, the next critical split is at fossil fuel costs, identifying a set of scenarios (path #2) which do achieve high wind and solar energy fractions under cheaper fossil fuels. This, again, is a less likely path toward high solar and wind energy, with only 44 out of 124 scenarios with these parameters ending up in the top 5%. CART identified two other parameters as key to high wind and solar under these conditions: constrained bioenergy and cheap solar storage. Lower fossil fuel costs can make fossil fuel electricity generation combined with carbon capture and storage technology a more economically compelling option for meeting climate targets than renewables, making this perhaps a less expected driver of high renewable energy consumption. However, if bioenergy is constrained and solar storage is cheap, in addition to the cheap wind and solar technology and expensive carbon capture and storage, 28 scenarios end up with high wind and solar energy fractions.
The combination of low wind and solar technology costs, high CCS costs, high population and GDP, and high fossil fuel costs was the most reliable path to high wind and solar electricity in our analysis (104 of 128 scenarios). This set of conditions is further divided into two paths. Path #3 is further defined by unconstrained bioenergy, a factor expected to reduce wind and solar somewhat, and low solar storage costs, and makes up 16% of the top 5% of scenarios. By far the majority of the top 5% of scenarios are in Path #4 (32%) with high population and GDP, high fossil fuel costs, and limited bioenergy, all of which we would expect to contribute to higher solar and wind consumption.
The different paths to achieve high levels of wind and solar energy represent different types of future worlds and come with synergies and tradeoffs in terms of other modeled outcomes that hold relevance to human well-being and other societal goals ( Figure 4). Generally, reaching high levels of renewable energy when bioenergy is unrestricted (Path #3) is associated with higher food prices across nearly all regions and higher air pollutant emissions in many regions. The former is due to competition for land allocation between bioenergy and food crops (consistent with results seen in Dolan et al., 2022). The latter are largely driven by changes in unmanaged land area that affect grassland and forest burning, as well as directly from biomass combustion. On the other hand, reaching high fractions of wind and solar electricity under a low population and GDP future (Path #1) leads to clear synergies including lower food prices, air pollutant emissions, and water consumption. Under more expensive fossil fuels and limited bioenergy (Path #4), the scenarios with high fractions of wind and solar energy also generally had higher fractions of nuclear energy. The lower fossil fuel prices path (Path #2) was generally more middle-of-theroad in terms of outcomes, with few significant differences between its mean values and those of other paths.
Beyond these more broadly consistent characterizations of each path, we found a wide range of variability between regions within each of the different paths. We focus on water consumption and carbon monoxide emissions, each in three different regions, as illustrative examples to further explore some of this regional heterogeneity ( Figure 5).
For water consumption we compare Brazil, Japan, and the United States, with these particular regions chosen for illustrative purposes as each region has different water consumption outcomes within each path and highlights multi sector tradeoffs that can occur in different futures defined by the same high level criteria.
Water consumption across all three countries was low under low population and GDP conditions (Path #1), but driven by different factors in each country: reductions in sugar crop production in Brazil, lower municipal Figure 4. Regional synergies (blue) and tradeoffs (red) of different paths to high combinations of wind and solar electricity generation in Global Change Analysis Model. A set of eight example regions were selected to illustrate regional variability, two from each of four levels of gross national income (GNI) per capita ( Figure S4 in Supporting Information S1). The outline type on each bar corresponds to statistical significance, where a dashed line is not a statistically significant result.
water use in Japan, and reductions in water used for corn irrigation in the United States. With high wind and solar electricity generation under an unrestricted bioenergy path (Path #3) in Brazil, a substantial drop in sugar crop water consumption leads to an overall lower water consumption in this path compared to the other three.
In Japan and the United States, by contrast, the increase in biomass water consumption increases overall water consumption with higher bioenergy levels (Path #3, Figure 5a). Changes in corn production in the United States is the second strongest driver in this path, and is the dominant driver of changes in water consumption across the other three paths.
For carbon monoxide (CO) emissions, a harmful air pollutant with serious health impacts, we selected China, Brazil, and Western Africa, again for illustrative purposes to explore regional heterogeneity in trends and sectoral drivers. In Western Africa, an increase in residential sector emissions drove higher CO emissions under low population and GDP (Path #1). All three other paths in Western Africa lead to overall reductions in carbon monoxide emissions, thanks to reductions in residential and unmanaged land emissions. In Brazil, changes to unmanaged land area drive changes in carbon monoxide emissions across all four paths due to biomass burning and forest fires. This leads to high CO emissions under an unconstrained bioenergy path (Path #3), and lower emissions under lower population and GDP conditions (Path #1). Drivers of CO emissions in China are somewhat more varied, with electricity sector emissions playing a major role, along with transportation and unmanaged land. Under low fossil fuel prices (Path #2), higher electricity sector emissions lead to an overall increase in CO emissions, while under a low population and GDP path (Path #1), declines in both electricity and transportation sector emissions lead to an overall reduction in CO emissions (Figure 5b).

Discussion and Conclusions
Solar and wind energy levels in our ensemble are largely consistent with other results in the broader literature examining scenarios with similar warming targets. The fractions of wind and solar electricity generation in our ensemble ranged from 35% to 63% in 2050, a narrower range than that seen in the IIASA1.5 scenario database, which had fractions of combined wind and solar electricity generation of 16%-92% (Huppmann et al., 2019). The IIASA1.5 database has scenarios that span a wider range of emissions and temperature outcomes than ours, so fitting within but not fully spanning this range is in line with expectations.
Our results support the importance of solar and wind technology cost reductions, along with reductions in solar storage costs, for achieving high levels of wind and solar energy in a variety of different future worlds, each with different tradeoffs across sectoral outcomes. Wind storage costs, on the other hand, had low relevance in every region due to the lower variability of wind compared to solar power generation, and thus reduced reliance on storage. Additionally, this ensemble analysis highlights the relevance of carbon capture and storage costs, since cheaper CCS enables meeting scenario emissions targets through fossil and bioenergy, combined with CCS technology, rather than solar and wind. Population and GDP levels and fossil fuel resource costs were secondarily important.
Somewhat counter to expectations, several parameters played little to no role in controlling the top fractions of electricity generation in any region: backup electricity generation requirements, nuclear costs, emissions constraints, electrification, and wind storage costs; however, this does not mean these parameters have no importance in the real world future. A more comprehensively sampled parameter space may have different importance attributed to these parameters. Additionally, the limited role of these parameters may be influenced by model structure, and further analysis with other multi-sector models will be useful to explore the influence of this uncertainty on our findings.
Nonetheless the ensemble considered in this study is still sufficiently large to highlight the complex interactions and trade offs across sectoral metrics in different regions, even when the ensemble is developed with a specific outcome in mind. These findings of parameter importance were generally consistent across different choices for the "top" classification (4%, 5%, 6%, or 7%) level ( Figure S5 in Supporting Information S1), though some parameters such as wind technology costs, fossil fuel costs, and CCS costs changed rank more substantially across the different "top" classifications.
In this work we focused on two emissions scenarios with low temperature targets, which limited the influence of carbon taxes on solar and wind penetration, in contrast to some previous analyses (e.g., Jaxa-Rozen & Trutnevyte, 2021). GCAM has a variety of low carbon technologies, so while carbon prices affect the demand for low carbon energy, the model does not prioritize wind and solar over other low or zero emissions technologies as a result of carbon prices. Electrification has also been previously found to be significant in increasing renewable energy (Luderer et al., 2022), and while in this ensemble we do see a stronger influence of electrification on absolute values of wind and solar electricity generation, we focus this analysis on renewable shares, which limits the influence of higher or lower overall reliance on electricity.
This ensemble stayed away from exploring questions of the structure of solar and wind energy modeling in GCAM, instead focusing primarily on cost parameters with only a few exceptions. For example, uncertainty underlying the fundamental supply curves for wind and solar in GCAM was not explored here and has been previously found to be influential on future wind and solar energy distributions (Silva et al., 2021). We also limited our focus to the top 5% of wind and solar scenarios at a global scale, which were not identical to the top scenarios in different regions, so regional differences between paths to high renewables in a specific region may vary from the results discussed here.
Our ensemble results span a wide range of energy and technology parameters within GCAM and highlight key drivers and combinations of factors that affect mid-century solar and wind energy consumption. Globally and regionally, solar and wind-related technology costs were primary drivers, though a few regions depend heavily on other parameters like carbon capture and storage costs, population and GDP trajectories, and fossil fuel costs. These results additionally demonstrate how scenarios with specific characteristics at a global scale (high solar and wind energy adoption) have a variety of different impacts at the regional scale.
In the large ensemble of scenarios explored, which focused on two low temperature target scenarios, we found that the costs of renewable technologies were far stronger drivers of wind and solar adoption than the rate at which the emissions target was reached. Globally, high shares of solar and wind energy can be reached in several different ways, with some of the lowest air pollution, water consumption, and food prices in many regions associated with the low population and GDP path, and higher outcomes in these sectors associated with the unconstrained bioenergy path.