This week’s Tuesday Talk was given by our own Professor Philip Stier. Alongside being a Fellow at Reuben, he is Professor of Atmospheric Physics at the University of Oxford and Director of the Intelligent Earth UKRI AI CDT in AI for the Environment.
Weather, AI, and climate modelling
The talk began with an explanation of how traditional weather forecasting works. These forecasts use physics-based models to predict short- to medium-term weather trends. AI models are trained on these models and can now outperform them in many settings.
Prof Stier then moved on to discuss climate modelling. In theory, this is just using the same weather model but allowing it to run for much longer than the standard 14-day weather forecast. But, as the timescale increases, so does the uncertainty. Clouds represent one of the most significant uncertainties in both climate and weather forecasting. Clouds absorb and scatter light. Because they cover 70% of the Earth, they can have enormous effects on weather and climate.
Clouds and the climate
Something I’d never considered is that air pollution can form clouds and reduce the effects of global warming. Prof Stier showed this with a picture of a cloud forming above a shipping lane due to emissions from the ships below. This highlights that air pollution is a double-edged sword - it is hugely damaging to human health, but can reduce global warming.
The talk ended with a summary of current AI models for climate prediction. Interestingly, unlike weather forecasting models, big commercial companies are developing many of them. Currently, predictions from different models aren’t particularly consistent. Prof Stier argued that this is because they need to accurately model clouds!
Some interesting questions to explore
This presentation sparked an extensive and interesting Q&A session. One discussion looked at the climate costs of using AI itself to model climate change. Once trained, AI models are more energy efficient than traditional, physics-based ones. However, training them takes large amounts of energy, and the models keep growing to become more effective.
Another discussion considered the need for AI in climate modelling at all. We already have a scientific consensus that climate change is happening - will these predictions change anything?
Several questions focused on the role of private companies in climate modelling, which can lead to a power imbalance. Though companies describe systems as ‘open source’, they often don’t release their training pipeline. This discussion also highlighted the need for ways to benchmark models. However, it can be especially hard when the models are predicting so far into the future.
Overall, the evening was enjoyable and informative, and accompanied by some excellent food!
