The Knowledge Frontier

Earth & climatePreprintData analysis3 min read

HOW MUCH OF RECENT CLIMATE CHANGE IS NATURE'S OWN SWING?

Every climate trend we observe is a sum of two things. One is the forced response: the change pushed by greenhouse gases, aerosols, land use and volcanic eruptions. The other is internal variability: the climate’s own swings, such as El Niño or the slow oscillations of the Pacific and Atlantic oceans. Telling them apart matters, both to understand climate change and to judge whether climate models get it right.

The trouble is that reality runs only once. Climate models can be run dozens of times with the same forcing, and the average of those runs reveals the forced part. But models have biases, and simpler statistical tricks tend to assume that the forced pattern never changes. A recent international comparison of such methods found that errors remained large — and that machine-learning methods did no better than classical statistics.

The paper offers a cautionary example: the slowdown of global warming between 1998 and 2012 was first read as a sign that human-driven warming was slowing and that models were wrong, before being recognised as internal variability.

A network that learns the noise

Parvathi Kooloth, Wenyu Zhou and colleagues, at the Pacific Northwest National Laboratory and the universities of California San Diego and Washington, built a ConvLSTM, a neural network that combines a sense of spatial patterns with a memory of how they evolve. They trained it on large ensembles of simulations from 15 climate models, using maps of surface temperature, rainfall and sea-level pressure. Its task was to recognise the natural swings; whatever is left is the forced part.

Tested on models it had never seen, it beat six existing methods for temperature, rainfall and pressure over most of the historical period — with one exception, a simulation the authors argue is an unrealistic outlier. For the real world, ten versions of the network, each trained on a different set of ten models, were applied to the observed climate since 1940.

What it finds in the real record

  • The warming slowdown of 1998–2012 comes out as natural variability, as later studies concluded. The natural part of global temperature follows the Pacific’s slow oscillation closely.
  • The Arctic. Observed Arctic warming has recently reached about four times the global rate. According to the network, the forced part stays almost constant at about three times. Natural swings did the rest: the Pacific held global warming back while a regional back-and-forth of its own, an Arctic mode of variability, boosted warming there, especially in 2005–2020.
  • The Atlantic. The temperature contrast between the North and South Atlantic has followed a V shape since 1950. The forced part also dips and rises, in line with the rise and later decline of aerosol pollution in the Northern Hemisphere. Natural variability amplified the swing, creating the deep low of the 1970s and 1980s.
  • The Pacific and the US Southwest. Since 1980, the eastern equatorial Pacific has warmed less than the west, a “La Niña-like” pattern that models rarely reproduce. The network says the forced response alone would have produced a weak opposite, “El Niño-like” pattern; natural variability reversed it. The same holds for the strengthening of high pressure over the northeast Pacific and the drying of southwestern North America: mostly natural, with only a modest forced share.

For several of these features, the trend observed over 1980–2022 lies outside the range of the estimated forced response — a reminder that a single historical trend can be a poor guide to what humans caused.

A check on the models

Because the network gives an explicit estimate of the forced part, it can be compared directly with what models simulate. The broad patterns agree: more warming over land and at high latitudes, less over the subpolar North Atlantic and the Southern Ocean, wetter tropics and high latitudes, drier subtropics. But the models warm the Northern Hemisphere more, strengthen subtropical high pressure less, and dry the land less. The authors suggest that current models may underestimate how much shifting winds will dry the continents.

They list the limits themselves. The network learns from models, so a bias shared by all the models could mislead it. Why it works so well is hard to see inside it. And it assumes that the average of a finite set of simulations is the true forced response. Within those limits, the authors conclude, machine learning can make the attribution of past climate change sturdier — and the testing of models sharper.