Climate Model Emulators

What are they, and why work on them?

Climate models are state-of-the-art tools for understanding how the Earth system works.  They’re also extraordinarily complex, computationally expensive models, which makes some aspects of science difficult, such as exploring different scenarios of climate change or understanding impacts.  Climate model emulators are simple versions of climate models that only represent part of the system or represent the system at reduced complexity.  By training and calibrating these emulators to represent the more complex Earth system models, we can get rapid climate information, often in a few seconds on a laptop.

What do we do?

A demonstration of machine learning trained on climate models and used to provide data-driven climatological “forecasts." From Weber et al. (2019),

We build emulators of climate models and apply them to studies of scenarios and impacts.

  • Calibration. One can train emulators on different Earth system models to get them to represent those models.  Then we can emulate those models in different scenarios to explore the importance of model spread and uncertainty.
  • Machine learning. By training machine learning algorithms on climate model output, we can then generate new scenarios with realistic weather patterns and climate variability, which allows us to dive into details of things like extreme events.
  • Geoengineering. By training an emulator on climate model output representing climate change and geoengineering, we can explore different scenarios of geoengineering to understand ranges of model spread.  This is important for understanding a range of options for climate change:  for example, what are different ways in which society can limit global mean temperature rise to 1.5°C, and what are the consequences of those different scenarios?