A city might cut one air pollutant even as emissions of another continue to climb. That seemingly ordinary policy choice poses a problem for a climate model trained on futures in which the two tend to move together. In a study reported by MIT on November 12, an emulator trained on deliberately unusual emissions histories captured the effects of a local pollution-control scenario that the comparison models missed.
An emulator stands in for a more demanding physical simulation. Once trained, it can evaluate a scenario in seconds. The bottleneck comes earlier, when scientists generate the examples from which it learns. Running a detailed Earth system model for even one possible emissions trajectory can occupy a computing cluster for months.
The researchers used a simpler climate model to develop a different way of choosing those examples. They measured the error in the emulator and adjusted the synthetic emissions history to improve how well it learned. A sharp change in the output of a single pollutant can reveal a response obscured in a more plausible history. A test with scenarios generated by a full Earth system model still lies ahead.
What an unusual scenario can tell a planner
For a planner comparing two emissions policies, the important number may be the difference between two futures that otherwise look much alike. A model can track the broad warming trend in both and still get that difference wrong. Testing whether it recovers the effect of changing one pollutant gives a more direct measure of its value for that particular decision.