Machine learning algorithms are highly effective at modelling complex systems. But what about when that system is the whole of Earth? That's the question the Artificial Intelligence for Earth System Predictability (AI4ESP) initiative hopes to answer.
As Earth's average temperature rises, we see more extreme events (wildfires, dangerous heat, drought, hurricanes, flooding, and many others) which threaten lives, homes, and ecosystems. The better we can predict the impacts of climate change, the greater our ability to adapt to them - but this is easier said than done. Even with the advances in computing over the past decade, making predictions over microbe-to-global spatial scales and minutes-to-centuries temporal scales and capturing the non-linear interactions between humans and our climate remain challenging.
In a recently released report, AI4ESP workshop participants discuss the outlook for machine learning in climate prediction. The report points to the potential for AI calibration to replace manual tuning of climate models (a task that often takes years to complete). Machine learning could also be used to build technology that quality-checks data in near real time, and to develop "self-guiding" autonomous measurement devices.
The authors also suggest that, as well as being trained to label data based on human descriptors, AI could be used to identify climate signals not yet recognised by humans.
The authors note that a number of challenges remain. For example, many machine learning algorithms trained to replicate a particular dataset are biased towards predictions of mean values. To adapt to the worst effects of climate change, we need to be able to predict the extremes. Progress in these areas may come from altering the ways in which the models learn: for example, modifying the loss functions used for training, or using generative adversarial techniques.
With one AI-based study predicting that Earth may pass 1.5 degrees of warming in as little as a decade, we may need those AI-enhanced predictions sooner than we thought.