Researchers' Zone:
Is AI our new oracle predicting how fast the ice is melting?
Current ice sheet melting models are slow, inaccurate and expensive. The picture of the future is uncertain - but AI can change that.
We scientists keep a close eye on the ice in Greenland and the Arctic, because even though there is a lot of ice, you probably already know that the ice is melting faster and faster. This is due to global warming, which is emphasised by the latest IPCC report and calculations from GEUS.
But how fast is the ice melting and how much meltwater is actually being produced? We don't know for sure.
The models we use to predict how much ice will remain in the future and how quickly it will disappear are slow, complex and disagree on the amount of melt.
This is why we are now working on a new project to develop Artificial Intelligence (AI) models that can create faster surface mass balance models.
Out of balance
The Greenland ice sheet is constantly changing.
You can compare it to your bank account, money comes from some places and comes out to others.
Similar to your bank account, the mass of the Greenland ice sheet can increase and decrease.
At the surface the ice sheet grows when it snows, but it thins and retreats when the surface melts.
The surface mass balance describes this increase and decrease of snow and ice at the surface only.
On the Greenland ice sheet, snow falls most of the year, but in the summer months when the temperatures are high (sometimes 0 degree in the middle of the ice sheet and over 10 degrees at the margin), the surface at lower elevations melts and the water might run off the ice sheet.
In an ideal world, the winter mass increase would equal the summer loss.
But with global warming, the increase in surface mass in winter does not match the increased loss of surface mass in summer.
Where glaciers retreat, they sometimes leave behind lakes that are dammed by the old moraine. A lake in South Greenland still has snow and ice in early summer when the fjord in the distance is green. The lake is high up, and with AI we can account for the fact that there is less melting at higher altitudes (Video: Clément Cherblanc).
Our predictions have improved/have become more and more accurate
We, climate researchers, currently observe physical phenomena such as melting ice and attempt to describe them with equations that computers can solve.
This approach is called ‘numerical modelling’.
Those models create a coarse grid, and each point of the grid is supposed to represent what happens at that location on the ice sheet and nearby.
Those models work great if the grid is very fine, but the finer the grid, the longer the computation time.
Computational power has improved over the last decades, thus the spatial resolution (=how fine the grid is) of the models has increased.
In the figure below you can see an example of how our models have improved over time.
But the computing power is starting to fall short
However, the spatial resolution needed to have reliable predictions of the surface mass balance is so fine that it would require extremely powerful computers, or running the models would take years.
But we do not have such powerful computers. So where to look for new solutions?
Artificial Intelligence (AI) has made rapid progress in the last years, and is applied in various fields, including language processing (like ChatGPT), image generation (Dall-E), but also in medical applications and mechanical engineering.
AI models also spark the interest of climate scientists – including us.
In climate modelling, AI models can be used to approximate the predictions of the numerical models without solving complex physical equations.
Thus we will be able to make new predictions at a fraction of the computation costs, as this study indicates, where the researchers managed to make instantaneous predictions, whereas the associated numerical model took multiple weeks on a supercomputer.
However, how much faster we can make the calculations depends on the specific application, but the time consuming part in AI is the training of the model, that also depends heavily on the available hardware.
Speeding up the process
In our projects, we are developing AI models designed to mimic the predictions of numerical models.
Since we can run an AI model much faster than a numerical model, we can make many predictions instead of one in the same amount of time.
Since a single simulation from one of the current standard models only produces one prediction, we have limited knowledge of the reliability of this prediction.
With the multiple outputs produced by the AI model, we get a range of possible predictions, that we can analyse together and get a more accurate picture of how likely the predictions are.
How to make an AI model mimic the old models (just faster)
AI models can be made in many ways, and each of those can also be build in different ways, with different ‘structures’.
This is why ChatGPT is not the same model as your Photoshop assistant, your phone fingerprint detector, or your phone’s camera object detector, or what we use for scientific predictions - those are different AI models with their respective strengths.
So before we can use the AI model to make predictions, we need to teach the models exactly how to mimic the numerical models.
We do that by giving the existing predictions from the numerical models to the AI model, so that it learns how the output should look when given a certain input. This we call training the AI model.
For AI models to be reliable, we need to train the model with a lot of data.
Gathering enough data to train a model can be tricky, especially when this data comes out of a numerical model that is very slow to run in the first place.
However, once the AI model is properly trained, we don’t have to run the numerical models again and thus reducing the computational cost.
AI models are only as good as their training data
One of the biggest challenges of AI models is that the models are only as good as the data they have been trained on.
Although we aim to correctly mimic the numerical models, they might not perfectly portray the real world and reliably predict the future, as we saw in the first figure the models make very different estimates of ice loss - even for the present-day ice sheet.
For one when we run the numerical models, we make a lot of compromises, which means we don’t capture the nitty bitty detail of what’s going on at the surface of the ice sheet.
Since we use the output of the numerical models to train the AI model, they inherit this.
We don’t know the future - and neither does AI
Another aspect of this challenge is that we can never fully know what the future holds.
If there are some fundamental changes about the way the ice sheet behaves in the future, we might not have included that in our equations yet, since it is extremely difficult to predict.
Tipping points are a good example of drastic changes in the system.
This article gives a good overview of different possible tipping points, and also shows that some elements that were proposed as tipping points but are no longer seen as such.
Because the traditional numerical models use these equations, they might not capture these future changes correctly.
And since the AI models are trained on the output of the numerical models, both the numerical models and the AI might not perfectly represent the future real world.
This challenge is not unique to AIs used to predict melting ice or climate change, this is the case with all AI models - they are only as good as the data they are trained on.
However, that doesn't mean we can't use AI to make predictions. As mentioned, AI models will both produce predictions much faster and in greater numbers.
This gives us the opportunity to compare predictions, calculate uncertainties and thus get a better picture of what the future holds for our ice caps.
