I am obsessed with “large initial-condition ensembles”—LICE. These are climate models that run multiple scenarios to reflect uncertainties about the system’s past. I’ve been working with them for the last year. Most major climate modeling centers have LICE. They are a valuable scientific tool for trying to understand the relative sizes of a human-caused global warming signal and the noise of natural climate variability.

LICE work like this. You take the same climate model and run it dozens of times, starting from a climate state in 1950 or earlier. In each run, the model is driven by exactly the same external factors—such as human-caused changes in atmospheric levels of greenhouse gases—but starts from slightly different “initial conditions” of the atmosphere and ocean.

One way of varying these initial conditions is by choosing the weather from different days. Another way of scrambling the initial state is by introducing a small random perturbation to the distribution of clouds. Because the climate system is complex and nonlinear, small differences in initial conditions grow over time. Within weeks, the atmosphere has little or no “memory” of its initial state. Within years to a few decades, the ocean, too, “forgets” the initial three-dimension ocean state.