Surface climate emulators

To make informed decisions about SAI, decision makers need a wider variety of studies that help them understand the potential impacts of SAI on specific policy-relevant metrics (e.g., water availability, regional crop yields, fisheries productivity, etc.). Running impact studies focused on surface climate responses can be time-consuming and disjointed. Developing a small set of emulators (machine learning models trained to approximate the outputs of full climate simulations without needing to rerun the simulations each time) would allow researchers to evaluate the impacts of various SAI injection strategies and scenarios on a wide range of surface climate responses more rapidly. Supporting the development of these tools would accelerate research across the various activities related to individual climate uncertainties.

This activity will develop machine learning-based emulators that rapidly predict surface climate responses to different SAI injection strategies without requiring full Earth System Model (ESM) simulations. The emulator will be trained on existing GeoMIP simulations, other Earth system model outputs, or observations or reanalyses, and will learn the relationships between injection parameters (location, magnitude, timing, scenario) and resulting surface climate variables (temperature, precipitation, sea level pressure, etc.). A second category of emulators will extend this to include non-climate variables (impacts), such as agricultural responses or fire risk.

Such emulators are beginning to be built (Farley et al., 2026; Beall et al., 2025), but further development includes expanding the training sets to cover more injection strategy and scenario space, incorporating additional models and validation, using more sophisticated emulation approaches, and extending emulation to a broader range of surface climate (or whole atmosphere) responses and to climate impacts.

Development will include:

1. Building the training dataset — compiling and preprocessing output from GeoMIP simulations to create a structured training dataset with injection parameters as inputs and surface climate responses as outputs.
2. Designing and training the models — designing and training machine learning models capable of providing both point predictions and uncertainty quantification reflecting multi-model spread.
3. Validation — validating the emulators.

The final deliverable will be publicly available emulator tools released with comprehensive documentation and example notebooks. This will enable other researchers to use the tools to rapidly sample the scenario space, for example to conduct additional climate impact assessments without running new ESM simulations, as well as enabling those outside the scientific field to explore the consequences of different SAI strategies quickly and interactively.