El Nino and La Nina arise from interactions between the tropical Pacific Ocean and atmosphere. Climate models generally expect this cycle to change as the planet warms, but they disagree about how much its sea-surface-temperature variability will change. ^[1]

One reason is that some models do not reproduce the observed relationship between Pacific warming patterns and El Nino variability. The researchers asked whether a neural network could use historical observations to identify the model relationships that look more like the real ocean. ^[1]

The team trained 11 neural networks on simulations from 11 CMIP6 climate models. Each network learned a relationship between tropical Pacific warming patterns and the amplitude of El Nino sea-surface-temperature variability. The networks were then evaluated with three observational datasets rather than accepted on simulation performance alone. ^[1]

Networks that matched observations tended to focus on physically meaningful regions in the central and far-eastern equatorial Pacific. This interpretability check does not prove the networks have learned every relevant process, but it helped the researchers reject models whose historical behavior was less plausible. ^[1]

The authors describe this change as a 54% reduction in uncertainty. It narrows the spread in one measure of El Nino variability under one emissions pathway; it does not predict the date, strength or impacts of a future El Nino event. ^[1]

The method uses climate-model simulations for training and only 11 models met the study's data requirements. Observational records are much shorter than the simulated climate histories and carry their own uncertainty. A model-as-truth test checks internal robustness, but it is not observation of the future. ^[1]