Communications of the ACM
Modelling Clouds and Climate
While it's still early, some work suggests incorporating machine learning cloud models into climate models does indeed improve predictions.
Credit: Varun Singh Bhati/EyeEm/Getty Images
Clouds have a significant effect on the Earth's climate. They can block the sun and have a cooling effect, or trap heat and add warmth, depending on factors such as their height, how reflective they are, and whether it is day or night. In addition, they produce rain.
Global climate models, however, which are used to predict future climate, typically do not do a good job at simulating the contribution of clouds. Their spatial resolution is typically 50 to 100 kilometers, whereas clouds can be as small as a few hundred meters. That's why detailed physical models of the small-scale processes that control clouds, such as convection in low-level clouds and deep convection in the tropics, are not included in climate models.
"The fact that we don't resolve (these processes) introduces a lot of variability and uncertainty in how much extreme precipitation events will change (in the future)," says Janni Yuval, a post-doctoral fellow in the Massachusetts Institute of Technology (MIT) Department of Earth, Atmospheric, and Planetary Sciences (EAPS).
The main limitation is computational power. Even the powerful supercomputers used to run climate models would need to run for years to obtain higher-resolution results, which is not feasible, since many research groups typically share the use of supercomputers.
Yuval and a growing number of researchers are looking at how machine learning could be used as a shortcut. Climate models currently include cloud processes to some extent by incorporating simplified representations of the physics involved, but machine learning would take a different approach: algorithms would learn how clouds contribute to climate directly from high-resolution data. Once that is accomplished, the algorithms could be plugged into a lower-resolution climate model.
"The fact that we can learn directly from data that we believe is very similar to how the physics of the system behaves gives us the ability to develop a data-driven approach instead of a simplified theory," says Yuval. "The advantage is that it will be much more accurate."
Although it's still early, some work suggests incorporating machine learning cloud models into climate models does indeed improve predictions. Yuval and his colleagues, for example, were able to better represent precipitation extremes in recent work.
"Using machine learning models produces precipitation extremes that are much more similar to the high-resolution model," says Tom Beucler, an assistant project scientist in atmospheric science at the University of California, Irvine, and Columbia University. "They are giving you the properties of a higher-resolution model 100 times faster."
However, there are concerns to address with this process as well. Machine learning systems are unaware of the laws of physics so they often produce results that break them, which could lead to inaccuracies in long-term climate projections. "Laws like conservation of mass and conservation of energy are typically violated by almost every machine learning algorithm," says Beucler.
Physical rules can be built into machine learning models, though. Beucler and his colleagues showed that techniques can be used to enforce conservation laws without compromising on accuracy when modelling cloud convection. Their system, which used neural networks, essentially was penalized when it violated physical laws, so it learned to avoid those cases. "The main innovation is that because the conservation laws are inside the neural network, it's going to feel the effect of the conservation laws as it is optimized, and adapt to them," says Beucler.
Interpretability is also a major issue. Since machine learning systems learn on their own, what they have learned is typically a black box.
"A core motivation of climate science is to deliver a long-term prediction to help countries make decisions on how to adapt and mitigate climate change," says Beucler. "If you have a model that's not interpretable, people are not going to feel comfortable going to policymakers and telling them what to do because they're not sure why that number came to be."
Since understanding results is important in many machine learning applications, computer scientists have been working on techniques that can help reveal what the algorithms are learning. Beucler and his colleagues used one such technique to show their machine learning model of storms and cloud convection is physically consistent. "It does not contradict what we know for sure about storms, and that makes the model more trustworthy, or less of a black box," he says.
If machine learning models of cloud processes are to be used more widely, they need to hold up when inserted back into climate models and, ideally, to be capable of working in several different ones. Models still often crash when plugged back in, though. Several factors may be at play: a machine learning model may simply not be incorporated correctly, for example, or it may not be able to handle the conditions of certain climate models.
Yuval and his colleagues, however, were able to develop a machine learning cloud model that works well with different climate models. Yuval thinks it is successful partly because it incorporates a lot of the physics involved. "It was a big breakthrough to show that it never crashes," says Yuval. "We could run it (in climate models with) multiple scenarios and resolutions and it works."
As machine learning models improve, they may be able to do more than just simulate cloud behavior. Beucler thinks they could lead to new discoveries, since they are able to make sense of complex behavior that is still not well understood.
"Once you have a machine learning model, you can try to extract a lot of knowledge from everything that it's learned from data," he says. "Once people realize (it can do that), I think it will really take off in this field."
Sandrine Ceurstemont is a freelance science writer based in London, U.K.
No entries found