Climate models are a key technology for predicting the impacts of climate change. By running simulations of Earth's climate, scientists and policymakers can estimate conditions such as sea level rise, flooding, and rising temperatures, and make decisions about how to respond appropriately. But current climate models struggle to provide this information quickly or affordably enough to be useful at smaller scales, such as the size of a city.
Now, the authors of a new open access document published in the Journal of advances in Earth systems modeling have found a method to leverage machine learning to utilize the benefits of current climate models, while reducing the computational costs required to run them.
“It turns conventional wisdom on its head,” says Sai Ravela, a senior researcher in MIT's Department of Earth, Atmospheric, and Planetary Sciences (EAPS), who wrote the paper with EAPS postdoc Anamitra Saha.
Traditional wisdom
In climate modeling, downscaling is the process of using a coarse-resolution global climate model to generate finer details in smaller regions. Imagine a digital image: A global model is a large image of the world with a small number of pixels. To scale down, you zoom in on just the section of the photo you want to view (for example, Boston). But because the original image was low resolution, the new version is blurry; does not provide enough detail to be particularly useful.
<p paraid="1398159223" paraeid="{739dfac8-6cd7-4eaf-869c-1e5024e4c61c}ethereum“>“If you go from a coarse resolution to a fine resolution, you have to add information in some way,” explains Saha. Downscaling attempts to add that information back by filling in the missing pixels. “That addition of information can happen in two ways: it can come from theory or from data.”
Conventional downscaling often involves using physics-based models (such as the rising, cooling and condensation process of air, or the landscape of the area) and supplementing them with statistical data taken from historical observations. But this method is computationally demanding: it requires a lot of time and computing power to execute, and at the same time it is expensive.
A little of both
In their new paper, Saha and Ravela have discovered a way to aggregate the data in another way. They have employed a machine learning technique called adversarial learning. It uses two machines: one generates data for our photo. But the other machine judges the sample by comparing it with real data. If it thinks the image is fake, the first machine must try again until it convinces the second. The ultimate goal of the process is to create super-resolution data.
The use of machine learning techniques such as adversarial learning is not a new idea in climate modeling; Where it currently has problems is its inability to handle large amounts of basic physics, such as conservation laws. The researchers found that simplifying physics and supplementing it with statistics from historical data was enough to generate the results they needed.
“If you augment machine learning with insights from statistics and simplified physics, it suddenly becomes magical,” Ravela says. He and Saha began by estimating extreme amounts of rainfall by eliminating more complex physical equations and focusing on water vapor and Earth's topography. They then generated general rainfall patterns for both mountainous Denver and flat Chicago, applying historical accounts to correct the results. “It is giving us extremes, as physics does, at a much lower cost. And it gives us speeds similar to those of statistics, but with a much higher resolution.”
Another unexpected benefit of the results was how little training information was needed. “The fact that just a little bit of physics and a little bit of statistics was enough to improve the performance of the ML (machine learning) model…wasn't really obvious from the beginning,” Saha says. It only takes a few hours to train and can produce results in minutes, an improvement over the months it takes other models to run.
Quickly quantify risk
Being able to run models quickly and frequently is a key requirement for stakeholders such as insurance companies and local policymakers. Ravela gives the example of Bangladesh: by seeing how extreme weather events will affect the country, decisions can be made about what crops should be grown or where populations should migrate to considering a very wide range of conditions and uncertainties as soon as possible.
<p paraid="1361710504" paraeid="{cf62482c-5732-4cf2-8218-94f262acd9bb}stocks“>”We cannot wait months or years to be able to quantify this risk,” he says. “You need to look into the future and a lot of uncertainties to be able to say what might be a good decision.”
While the current model only looks at extreme precipitation, the next step of the project is to train it to examine other critical events, such as tropical storms, winds and temperature. With a more robust model, Ravela hopes to apply it to other places like Boston and Puerto Rico as part of a Great Climate Challenges Project.
“We are very excited about both the methodology we have developed and the potential applications it could lead to,” he says.
