Machine learning accelerates climate model simulations at finer resolutions, making them usable at spatial scales.

Climate models are a key technology in predicting the effects of climate change. By running simulations of Earth's climate, scientists and policymakers can predict 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 cheaply enough to be useful at small scales, such as the size of a city.

Now, the authors of a new paper published in Journal of Advances in Modeling Earth Systems have found a way to leverage machine learning to harness the benefits of existing climate models, while reducing the computational costs required to run them.

“It turns conventional wisdom on its head,” says Sai Ravela, a principal research scientist in MIT's Department of Earth, Atmospheric and Planetary Sciences (EAPS), who is an EAPS postdoc in Anamitra. Wrote thesis with Saha.

Conventional wisdom

In climate modeling, downscaling is the process of using a global climate model with a coarser resolution to produce finer detail over 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 part of the image you want to see — Boston, for example. But since the original image was low resolution, the new version is blurry. It doesn't give enough detail to be particularly useful.

“If you go from coarse resolution to fine resolution, you have to add information somehow,” explains Saha. Downscaling tries to add this information back in by filling in the missing pixels. “Information can be added in two ways: either it can come from theory, or it can come from data.”

Traditional downscaling often involves the use of models based on physics (such as wind upwelling, cooling and condensation processes, or the topography of an area) and supplementing it with statistical data taken from historical observations. to do But this method is computationally taxing: it takes a lot of time and computing power to run, while it's also expensive.

A bit of both

In their new paper, Saha and Ravela have found a way to incorporate the data in another way. They have used a technique in machine learning called adversarial learning. It uses two machines: one prepares the data to go into the image; But another machine makes a decision by comparing the sample with the actual data. If it thinks the image is fake, the first machine has to try again until it convinces the second machine. The ultimate goal of the process is to create super-resolution data.

Using machine learning techniques in climate modeling is not a new idea like adversarial learning. Where it currently struggles is its inability to handle large amounts of fundamental physics, such as conservation laws. The researchers discovered that simplifying the physics and supplementing it with statistics from historical data was enough to produce the results they wanted.

“If you augment machine learning with some information from both statistics and simple physics, all of a sudden, it's magical,” says Ravella.

He and Saha began estimating extreme precipitation amounts by removing the more complex physical equations and focusing on water vapor and land topography. They then developed generalized rainfall models for hilly Denver and flat Chicago alike, applying historical accounts to correct the output.

“It gives us, at a very low cost, the same as physics. And it's giving us the same speed as statistics, but at a much higher resolution,” Ravela continues.

Another unexpected benefit of the results was that very little training data was needed. “The reality is that just a little bit of physics and a little bit of statistics was enough to improve the performance of ML. [machine learning] The model … was not really obvious from the start,” says Saha. It takes only a few hours to train, and can produce results in minutes, an improvement over the months it takes to run other models.