Scientists at NOAA’s Geophysical Fluid Dynamics Laboratory are Tapping into Machine Learning to Better Understand the Impacts of Climate Change on our Oceans and Atmosphere
When you hear the term “machine learning,” you might think of controversial chatbots or the algorithms that govern your social media feeds. But NOAA GFDL scientists are investigating how to use machine learning in another way: to improve climate, weather and other earth system models.
Unlike traditional climate models, which make predictions by simulating land, ocean and atmospheric processes, machine learning allows systems to “learn” from results of those simulations.
“In contrast to models that follow a set of explicit and pre-defined rules, machine learning aims towards building systems that can learn and infer such rules based on patterns in data,” NOAA GFDL scientist Maike Sonnewald and co-authors Christopher Irrgang, Niklas Boers, and Jan Saynisch-Wagner write in Carbon Brief. “As a result, a new line of climate research is emerging that aims to complement and extend the use of observations and climate models. The overall goal is to tackle persistent challenges of climate research and to improve projections for the future.”
Though machine learning is still in the experimental phase when it comes to its use in weather and climate models, scientists are optimistic about its potential.
“It is a very useful tool for any climate research,” GFDL scientist Hiroyuki Murakami, who is currently researching how machine learning can improve our understanding of extreme rainfall, says. “I believe that machine learning can enhance our understanding of the physical mechanisms of climate change and the predictability of future conditions.”
Here are three ways that scientists at GFDL are incorporating machine learning into their research.
Bettering our understanding of extreme precipitation in Japan
Figuring out which precipitation events are considered “extreme” can be tricky. “Extreme” precipitation doesn’t just mean intense precipitation; it can also mean events that are rare within a certain time period. Rare events can include precipitation that shows an unusual spatial distribution – rain falling across the whole of Japan or just Tokyo, for instance. However, according to Murakami, it is impossible for scientists “to recognize rare precipitation events if they do not accompany heavy precipitation.”
Japan, for instance, has seen more extreme precipitation over the past 40 years, but scientists weren’t sure how much of that trend was distinguishable from natural variability – and how much could be attributed to climate change.
Enter: machine learning. A machine learning technique called the “autoencoder” can detect rare precipitation events without any prior information about the events. Murakami used the autoencoder to observe and simulate daily precipitation in Japan to detect outlier events. He found that, according to both models and observations, there has been a significant increasing trend in outliers over the past 40 years – which indicates that climate change has played a role in these rare precipitation events in Japan. The models predict this trend will continue over the next three decades.
As Murakami’s study notes, using the autoencoder for anomaly detection is a “work in progress, and far from a fully solved area of machine learning.” But Murakami says that, so far, the process of using the autoencoder has been smooth, and he thinks the tool could be helpful in other scenarios as well.
“It is now time to apply the autoencoder technique to other regions such as the contiguous United States,” he said. “The autoencoder can also be applied to other hydroclimate variables such as daily maximum and/or minimum temperature.”
Reducing errors in climate models
Traditional climate models break the ocean, atmosphere, and land up into many grid points in order to predict future climate conditions. But features that are too small or complex to be explicitly calculated in the model – such as cumulus cloud convection – are parameterized (i.e. approximated).
“It is generally accepted that with traditional climate models, the more physical processes one can explicitly resolve, the better,” says Spencer Clark, scientist at the Allen Institute for Artificial Intelligence and NOAA GFDL. “Parameterizations for the same processes can vary from model to model and introduce uncertainty into model predictions. This is particularly true for quantities like rainfall.”
That uncertainty that comes from parameterization makes rainfall predictions a good candidate for improvement via machine learning. Clark and his fellow researchers have tested machine learning’s ability to correct errors made through the traditional parameterization process – and so far have seen promising results, not only in improving precipitation, but also land surface temperature across a range of climates.
Clark says that the machine learning field has greatly expanded in the three and a half years he’s been in it, and he expects that expansion trend will continue. Though there’s still lots to do before machine learning is used operationally in earth system models – including more testing and refinement, as well as better understanding of the machine learning processes as a whole – Clark is excited to see how the field progresses.
“Machine learning provides a promising new way to approach many of the challenging modeling problems in our field,” he says. “The opportunities are vast and it will be exciting to see how things develop in the future.”
Better ocean data for improved climate models
Understanding the ocean is key to understanding – and predicting – climate. Maike Sonnewald, oceanographer at Princeton and NOAA GFDL, is researching how machine learning can help scientists understand ocean dynamics, such as heat transport, to better inform climate models.
Machine learning, she says, “allows a deeper and more intuitive understanding of what’s happening in different areas of the ocean.”
Machine learning can also be used to see what’s going on below the surface of the ocean. Though satellites can provide excellent data on the ocean’s surface, a challenge in ocean research is that scientists have to physically go out to sea to understand what’s happening beneath the surface – and whether it’s with a ship or with data-gathering robots, it’s a pricey endeavor. And similarly for modeling scientists, sharing large amounts of data on model results is very difficult – particularly if it’s across borders.
“Oceanographers have been working for a long time to infer what’s beneath the surface, because it’s so hard and expensive. But doing this with accuracy has been one of the grand challenges of the field,” Sonnewald says. Machine learning has helped Sonnewald “map” characteristics beneath the surface, providing key data for weather and climate models, and has also allowed her to better compare different models and model runs.
“We’re really just starting to scratch the surface” of what machine learning can do in models, Sonnewald says. “That’s really exciting.”
Original source can be found here.
