AI emulators are already rivalling physical models in weather forecasting; when it comes to climate, they will need to explicitly represent forcings and withstand conditions that have never been observed.
A review published on arXiv by Tom Beucler, J. David Neelin, Hui Su, Chris Bretherton and nine co-authors takes stock of the transition of AI from weather forecasting to climate projection. Autoregressive models trained on reanalyses now equal dynamic models of nowcasting on a sub-seasonal scale, with well-calibrated sets and a reduced cost. But on the climate horizon, the challenge is changing: it is no longer a question of starting from an initial state, but of reliably reproducing the statistical response of the system to forcings (greenhouse gases, land use) that are often outside the observed domain.
Two safeguards
The authors propose two minimum requirements: that external forcings enter the model explicitly enough to be modified independently of each other, and that robustness be tested outside distribution, on extremes and on counterfactual trajectories. By comparing the announced throughput of AI emulators to those of physical models ported to GPUs, they note that AI saves time by advancing only useful variables, at the desired resolution and with longer time steps. AI downscaling could also partially replace explicit high resolution and make local hazard assessment cheap.
Read the full article on the original source
Go to the source →