Data to evaluate the regional impacts of stratospheric aerosol injection

Recently, there’s been a marked uptick in activity around solar radiation modification — a family of approaches that aims to reduce global warming by reflecting sunlight back into space. Much of that activity, including growing private investment, centers on one method in particular: stratospheric aerosol injection (SAI). As interest grows, so too does the urgency for research that supports public scrutiny and informed decision-making around if and how approaches to SAI evolve. We need to understand the ways SAI could affect people, plants, animals, and ecosystems, and how those impacts could vary across the globe. We’re releasing an openly available, globally consistent dataset of downscaled SAI scenarios to help researchers probe these questions — and the uncertainties associated with downscaling itself.

Global climate models (GCMs) provide a way to study the physical outcomes of solar radiation modification without the risks of outdoor experiments. They’ve been useful for characterizing how SAI could change things like global mean temperatures, temperature extremes, and precipitation regimes.¹1K Ricke et al. (2023) Hydrological consequences of solar geoengineering Annual Review of Earth and Planetary Sciences However, GCM outputs often can’t feed directly into the impact models used to study topics like water availability, crop productivity, or public health. Downscaling is a method that translates coarse GCM data into a form that can be used in these finer-scale models.

We’re releasing a dataset generated by applying two different downscaling approaches to outputs from two GCMs, each modeling future widespread deployment of SAI. The dataset and the codebase used to produce it are available under permissive, open source licenses — anyone can freely examine, reuse, or build upon them. We’re also providing a set of tools to help researchers access, subset, and work with the data efficiently. Together, we hope this release supports three goals: (1) enabling researchers across the globe to explore the regional impacts that matter to them, (2) providing a benchmark for independent downscaling efforts, and (3) laying the foundation for more systematic research into how downscaling itself introduces uncertainty into our understanding of SAI impacts.

Among potential climate interventions, SAI poses unique ethical challenges. A single actor could plausibly deploy it, with effects felt across the planet; consequences can’t be fully known in advance, but would depend heavily on the specifics of deployment; and no established governance structure exists to constrain or coordinate decision-making. And because SAI would mask warming without removing the greenhouse gases driving it, its availability could become an excuse to delay decarbonization, rather than to advance it. Finally, SAI would need to be constantly deployed for as long as CO₂ concentrations remain high; if stopped abruptly, it would lead to rapid warming, known as “termination shock.” Research in this space can’t stand apart from those realities. For that reason, we prioritized a deliberative approach to designing our research, with external review cycles throughout the project. For a full account of these engagements and how they shaped our work, see the white paper from our collaborators at the Alliance for Just Deliberation on Solar Geoengineering.

In this article, we explain our downscaling approach and summarize the contents of the release. We also explain the dataset's limitations and provide examples of the kinds of research we hope it will enable.

What is downscaling, and how can it support SAI research?

Downscaling refers to a family of methods used to transform coarse-resolution climate data (e.g., 1°x1°) into a higher-resolution form (e.g., 0.25°x0.25°) that is more suitable for local and regional analysis. At a high level, it relies on two sources of information: observations of the past climate at fine spatial scales, and climate model simulations of past and future climate at coarser scales. Downscaling algorithms use the relationships between these to produce high-resolution datasets that are consistent with large-scale climate model behavior.²2The approach described here is statistical downscaling. The main alternative, dynamical downscaling, instead uses coarse climate model outputs as boundary conditions to run higher-resolution regional climate models over a smaller spatial domain. Dynamical methods have several strengths, including that they are physics-based, do not assume stationarity, and ensure output is physically consistent. But they are far more computationally expensive, and were not considered for this project.

Downscaling addresses two problems impact modelers can encounter with raw GCM outputs, both of which create barriers to studying important SAI impacts.

First, many impacts respond to climate variability at finer spatial scales than GCMs resolve. Snowpack and water availability, for instance, are strongly affected by fine-scale variation in elevation that raw GCM output cannot capture — a limitation downscaling has long been used to address.³3A Wood et al. (2004) Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs Climatic Change ^,44R Wilby and C Dawson (2012) The Statistical DownScaling Model: Insights from one decade of application International Journal of Climatology Downscaling could help us research any process sensitive to local topography, including impacts to snowmelt, flood risk, or coastal ecosystems.

Second, GCMs carry biases in their weather statistics even over the historical period, which is a problem for impact models calibrated against observed weather data. Most agricultural models fall into this category, which makes it hard to study potential crop impacts under SAI. That’s a problem, because crop productivity is one of the societal impacts we're most concerned about under climate change, and SAI’s net effect on it is genuinely unresolved.⁵5J Proctor et al. (2018) Estimating global agricultural effects of geoengineering using volcanic eruptions Nature ^,66B Clark et al. (2025) Maize yield changes under sulfate aerosol climate intervention using three global gridded crop models Earth's Future A downscaling approach that includes a bias correction step addresses this mismatch, enabling research into not just agriculture, but any field whose impact models are calibrated on weather observations, including ecology and public health.

Challenges introduced by downscaling

Downscaling can unlock the ability to look at SAI impacts, but it also introduces a new layer of uncertainty. There is no single “correct” way to statistically downscale a dataset, and specific implementation choices can lead to meaningfully different results. As a result it’s essential to document downscaling methods clearly and carry awareness of its limitations into downstream impact analyses.

More fundamentally, the process of downscaling SAI simulations both changes what questions researchers can ask, and determines what answers they find. The typical way to get a handle on that uncertainty is to apply multiple downscaling methods to the same GCM simulations and compare the results. As regional SAI impact research ramps up, it will be important to evaluate results using multiple downscaled datasets to understand if the findings are robust, or whether they’re an artifact of the downscaling process itself.

What we’re releasing

We are releasing a globally consistent downscaled dataset, the full codebase used to produce it, and tooling to help researchers easily access the data. The downscaled data includes global projections of temperature, precipitation, and surface downwelling solar radiation at a 0.25° resolution, produced using two downscaling methods, under several future scenarios — an SAI deployment scenario, a termination shock scenario, and a non-SAI baseline — and the historical reference period.

We downscaled outputs from two global climate models — CESM2(WACCM6)⁷7A Gettelman et al. (2019) The Whole Atmosphere Community Climate Model Version 6 (WACCM6) Journal of Geophysical Research: Atmospheres ^,88G Danabasoglu et al. (2020) The Community Earth System Model Version 2 (CESM2) Journal of Advances in Modeling Earth Systems and UKESM1.1.⁹9M Mulcahy et al. (2023) UKESM1.1: Development and evaluation of an updated configuration of the UK Earth System Model Geoscientific Model Development Each had already been run for the historical period, under a middle-of-the road emissions pathway without SAI (SSP2-4.5), and for an idealized SAI deployment called G6-1.5K-SAI,¹⁰10W Lee et al. (2026) G6-1.5K-SAI and G6sulfur: Changes in impacts and uncertainty depending on stratospheric aerosol injection strategy in the Geoengineering Model Intercomparison Project Atmospheric Chemistry and Physics in which greenhouse gas emissions continue at SSP2-4.5 levels while sulfate aerosols are injected into the stratosphere to hold warming to 1.5 °C.¹¹11To date, four GCMs have run the G6-1.5K-SAI experiment, but we were only able to get model output for all the necessary simulations to downscale two of these four GCMs. We also ran a new termination shock scenario using CESM2(WACCM6), in which the SAI deployment represented in G6-1.5K-SAI halts abruptly,¹²12C Zarakas (2026) Climate model output from simulation of abrupt termination of stratospheric aerosol injection Zenodo following the Geoengineering Model Intercomparison (GeoMIP) protocol.¹³13D Visioni et al. (2026) The Geoengineering Model Intercomparison Project (GeoMIP) contribution to CMIP7 – description of new experimental protocols and preliminary results Geoscientific Model Development For each model-scenario combination, we downscaled all available ensemble members.

We processed the raw GCM data into a common format, including standardizing variable names and units (a step known as “CMORizing”) and documenting how different simulations and ensemble members branched from each other. We then downscaled the data using two algorithms: a daily variant¹⁴14B Thrasher et al. (2012) Technical note: Bias correcting climate model simulated daily temperature extremes with quantile mapping Hydrology and Earth System Sciences of the well-established¹⁵15For example, the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) and the World Bank downscaled datasets were both produced with this method and are widely used for impact studies. Bias-Correction Spatial Disaggregation (BCSD)¹⁶16A Wood et al. (2002) Long-range experimental hydrologic forecasting for the eastern United States Journal of Geophysical Research: Atmospheres ^,33A Wood et al. (2004) Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs Climatic Change method, and a bespoke method we refer to as Quantile Delta Mapping Spatial Disaggregation (QDMSD).¹⁷17QDM is an established name for the bias-correction step in the implemented downscaling method, but the full method has no standard name. We call it QDMSD, a term used elsewhere for similar (but not identical) methods. It uses the identical spatial disaggregation implementation as BCSD. Both were trained using the ERA5 reanalysis product.¹⁸18H Hersbach et al. (2020) The ERA5 global reanalysis Quarterly Journal of the Royal Meteorological Society A detailed description of our methods can be found here, and the full codebase is available on GitHub.

Including multiple downscaling methods in the output dataset allows researchers to start probing downscaling itself as a source of uncertainty. Both methods take the same approach to spatial disaggregation, but use different bias correction approaches. BCSD uses quantile mapping, which is not guaranteed to preserve climate model trends and can distort changes in extremes more than other downscaling methods.¹⁹19A Cannon et al. (2015) Bias correction of GCM precipitation by quantile mapping: How well do methods preserve changes in quantiles and extremes? Journal of Climate QDMSD uses quantile delta mapping,¹⁹19A Cannon et al. (2015) Bias correction of GCM precipitation by quantile mapping: How well do methods preserve changes in quantiles and extremes? Journal of Climate which is less likely to distort extremes, but can produce unrealistic values when applied to small initial values.²⁰20When quantile delta mapping (QDM) preserves trends as percent changes, rather than absolute differences, debiasing can generate unrealistically high values when dividing by small initial values. In the dataset, we include a set of embedded quality assurance variables, including flags where these problems occur. While the difference between these two methods may seem minor, it can influence the estimated effect of SAI relative to the baseline scenario (Figure 4).

The standardized GCM input data, the coarse bias-corrected GCM data, and the full downscaled dataset are published on Source Cooperative. All datasets are free to access on the cloud or to download. To support broad use, we are also releasing Jupyter notebooks designed to help researchers explore, customize, and download the data. The notebooks show users how to: subset by model, scenario, ensemble member, or variable; clip to a chosen geography; resample in time; apply common variable transformations; and export customized outputs in a range of file formats. See our docs site for more complete guidance on accessing and using the released data.

Limitations of the dataset

While we hope this dataset will enable new research into SAI impacts, it has limitations that are critical to understand. Each choice in the chain leading to a downscaled dataset introduces uncertainties — only some of which are visible in the structure of the data itself. Table 2 provides some guidance on how to think about each layer of uncertainty as you consider using the downscaled data.

We provide data at a fine temporal and spatial resolution because that is what many climate impact models need to run their experiments. But researchers should be careful not to allow the fine resolution to create a false sense of precision. Taken together, the choices outlined above mean the downscaled SAI dataset we’re releasing enables structured exploration of uncertainty — but is far from a complete characterization of it.

Because each downscaling step adds new assumptions and uncertainties, we encourage researchers to use climate data that is as close to the raw climate model output as their research question allows. Many important questions don’t need an impact model at all. For example, work on SAI’s effects on precipitation regimes or solar radiation can and should use raw GCM outputs directly. Others, like some vector-borne disease modeling, rely on impact models, but can be conducted at coarse resolutions. In these cases, the bias-corrected, coarse resolution data may be the better choice: Downscaling adds spatial detail that the analysis doesn’t strictly need, while adding uncertainty. The downscaled dataset itself is most appropriate for analysis that genuinely requires a high spatial resolution — like characterizing climate impacts for small island nations, or changes in mountain snowpack.

Research process

Although the dataset is the most visible output of this project, we see the process used to produce it as equally important. An accompanying white paper discusses our approach to research governance in full, but there are three aspects of the process that we want to highlight here.

01 — Project design and community input

This project began by gathering community input on both research purpose and design. During scoping, we held a workshop with 88 researchers at the 2025 GeoMIP conference. After committing to the project, we ran structured discussions about our project design with 33 individuals across 27 institutions and nine countries, spanning civil society, social science, and the downscaling, climate modeling, and impact modeling communities. Together, these early convenings informed key decisions, including the choice of downscaling algorithms and addition of a termination shock scenario. Near the end of the project, we invited 13 researchers to provide feedback on the data access utilities, and an initial version of the downscaled dataset itself. You can read a detailed summary of the convenings, what we heard, and how it affected project implementation here.

02 — Adding the termination shock simulation

One takeaway from convenings was the importance of downscaling multiple SAI scenarios, including suboptimal deployments, to capture the uncertainty related to scenario design itself. A commonly discussed example was a scaled deployment that stops abruptly. In this case, global temperatures would rapidly return to levels that would have occurred in the absence of the SAI intervention, placing huge strain on human and ecological systems.²¹21C Trisos et al. (2018) Potentially dangerous consequences for biodiversity of solar geoengineering implementation and termination Nature Ecology & Evolution This risk — known as “termination shock” — has been widely discussed, but was absent in the latest geoengineering model intercomparison simulations. Rather than work around that absence, we ran the simulation ourselves. We downscaled the output alongside the other scenarios, and are releasing the raw simulation output as well, so it can be used independently of the downscaled product.

03 — Transparency

The stakes around solar radiation modification are high. They demand research that serves the public interest, conducted with trustworthy methods and communicated clearly. Transparency is central to both. To that end, we are disclosing dedicated funding sources (the Bernard and Anne Spitzer Charitable Trust and The Navigation Fund), openly documenting the engagement process that shaped the project — including where we were and weren't able to incorporate community input, and why — and releasing all data and code under maximally permissive licenses. We hope these practices enable scrutiny and feedback, allow anyone to replicate and extend this work, and contribute to a culture of trustworthy research.

Moving forward

We’re excited to see what new research this dataset enables. Beyond the impact areas mentioned above — including crop productivity, changing water resources, and ecology — there are several other technical questions we’d be excited to see explored.

One is detectability: For example, how quickly would changes to water resources under SAI become distinguishable from internal climate variability, and how might that vary from place to place? Public perception of SAI would depend on observed regional climate changes, but in the decade following deployment, those regional signals would be driven largely by internal climate variability. Because we downscaled multiple ensemble members, researchers can separate the SAI signal from internal variability, making it possible to estimate how long SAI's effects would take to detect, and to begin unpacking how variability might shape public perception in the meantime.

We are also curious to understand how much uncertainty in projected SAI impacts stems from the choice of downscaling method itself. We hope that this dataset will prompt further research into downscaling as a source of uncertainty in SAI impacts research. While this is now possible with our dataset, there is a diverse ecosystem of downscaling approaches, and we expect that the two methods we implemented are a small sample of the true uncertainty. We hope that our codebase makes it easier to implement and assess a broader range of downscaling approaches, including regional dynamical downscaling.

As research into SRM continues, we anticipate a growth in regional impacts assessment using downscaled simulations. We hope that the dataset released here helps underscore some of the challenges of looking at regional impacts of SAI with downscaled data, and that our codebase will help researchers confront these challenges with greater nuance.