Tens of thousands of ships navigate the ocean everyday, leaving trails of emissions behind them that scientists have discovered can affect the formation of clouds. It’s something that, until recently, has been hard to track and required the painstaking task of sifting through satellite imagery by eye. But today, thanks to computation techniques developed in recent years, satellite images can be whizzed through in the metaphorical blink of an eye.
Advances in Artificial Intelligence (AI) like these in the last decade, particularly in the subfield of Machine Learning (ML), have equipped climate scientists with new tools for helping to save the planet. Researchers and policy makers are hoping to accelerate this progress through a series of expert talks organised by the International Telecommunication Union (ITU) under the AI for Good programme, starting with a kick-off event later today.
Philip Stier, a professor from the University of Oxford, who leads its Climate Processes group and is a moderator at the event, said machine learning was proving particularly useful for statistical modelling of the climate. “In a climate model, you often have uncertain representations of things like clouds, and increasingly we look to machine-learning methods to find simpler representations of the models that are consistent with the data,” he told Geneva Solutions.
Stier was joined on the call by Dr Duncan Watson-Parris, a senior researcher in the group, who specialises in machine learning. “Because of the variability of the climate and the natural world, no two events look exactly the same,” he explains. Scientists may not necessarily know in advance what they seek from their analysis. “For me, machine learning is about using computers, and particularly the power of large-scale computers, to find signals in the data without explicitly being able to tell the computer what it is that we’re looking for.”
Climate science is an inherently challenging field of research. An analysis must take into account several complex and ever-changing parameters. On top of this, there is only one opportunity in the real world to test long-term predictions made by the models; after all, you cannot travel back in time and try again. So, can machine learning help determine if a crucial piece is missing from an analysis?
Stier believes climate science is well-understood, generally speaking, on the theoretical side. “The problem is less that we think there’s something fundamental missing in our equations,” he continues. “It’s more that we know that our representations of these equations on the coarse scales of hundreds of kilometres are often inadequate. And it’s very hard to find when a model has such a structural error.” Where machine learning comes to the rescue is by providing a shortcut for representing a climate model, when traditional computing would require exploring all the parameters using a supercomputer running for a week.
Older forms of machine learning, such as clustering techniques, were adopted by climate scientists a long time ago. What has changed dramatically in the last decades, according to Stier, is that the data-science techniques have grown in scale, making new tools available to his fellow climate scientists. Watson-Parris adds that the huge amount of data now available for training machine-learning algorithms, coupled with more expressive algorithms, has moved the field along rapidly.
But like other fields of research, climate science is not immune to problems of algorithmic bias. Watson-Parris noted that this was a concern, citing an example of an algorithm that was trained on images of clouds taken during the daytime then struggling to identify the same clouds in images taken at night. This is where researchers’ expertise comes in.
“Machine learning is a clever way of fitting [the models],” Stier remarks, “but without particular knowledge of the field you could be fooled. You can learn a lot from statistics, but you always need to bring in climate physicists with domain knowledge, otherwise you could slip into some interesting conclusions!”
Machine learning is helping to perform research that was once impossible. Take events known as “pockets of open cells”, in which small regions of so-called open-cell clouds form within larger closed-cell regions over the ocean. Previous research on this phenomenon relied on the analysis of as few as one to five events, and the conclusions drawn have been challenged by the systematic analysis of around 9000 such events using machine learning. “It’s unlikely to have as big an effect as previously claimed in the literature,” says Stier. “Quantitatively looking at satellite data will be a game changer for all of environmental sciences.”
Despite encouraging progress, the use of AI in climate science is in its early days. “We’re still at the stage where we’re exploring which questions are amenable to these tasks,” Watson-Parris adds. There also remains scepticism about relying on “black box” AI tools. “That doesn’t mean that we can’t move ahead as a field,” Stier adds.
The AI for Good seminars will provide a forum to bring together a global community invested in advancing climate science through AI. Stier and the other experts participating in the forum also intend to publish a white paper at the end of the series of talks proposing several ways that climate scientists can move forward in their use of AI.
“We want to gain momentum from the seminars,” says Stier. “We feel there is a huge amount of expertise in the climate community and there is a huge amount of expertise in the data community and in industry. The excitement of this series is to bring people together and do stuff we couldn’t do before.”