A shortcut needs useful examples

A climate emulator is a faster stand-in for a slower simulation. It learns from another model’s output, so researchers can explore more possible emissions paths without running the slower model for each one. But a stand-in trained on a handful of similar paths can struggle when it sees something different.

In a paper published in PNAS on November 9 and described by MIT on November 12, researchers tested whether changing the examples could help. Instead of selecting only plausible future emissions paths, they designed a scenario to expose the training model to a wider variety of changes.

Why an unrealistic path can be useful

The team used a simple climate model to test the emulator and adjust the training scenario where its predictions fell short. The resulting paths could include quick rises or falls in emissions. Those swings are inputs to a training exercise, not a forecast of what will happen.

In the test using the simple model, the authors say one optimized scenario trained a better emulator on unseen scenarios than six standard pathways did. They also found an improvement when they used optimized scenarios to drive an intermediate-complexity climate model. The idea is to give the emulator enough contrast to learn which changes follow from different inputs.

What remains untested

These comparisons measure how well an emulator reproduces model outputs. They do not show that it forecasts the real climate better, or that the same gain will hold for a full Earth system model. MIT identifies testing at that scale as a next step.

The practical question is how best to spend a limited simulation budget. For the models studied, a scenario chosen to teach the emulator more proved more useful than adding familiar paths. Whether that trade-off holds for more detailed simulations is still open.

Sources

  1. MIT News: New research could lead to faster, trustworthy AI models for climate predictionNovember 12 university account of training procedure and planned test with a full Earth system model.
  2. Womack et al.: Optimal scenario design for climate emulationNovember 9 paper; accessible abstract and significance statement support the simple-model and intermediate-model comparisons.