How machine learning helps us refine climate models to unprecedented detail
From movie suggestions to autonomous vehicles, machine learning has revolutionized modern life. Experts are now using it to help solve one of humanity’s biggest problems – climate change.
With machine learning, we can use our abundance of historical climate data and observations to improve predictions of the Earth’s future climate. And these forecasts will have a major role in reducing our climate impact in the years to come.
What is machine learning?
Machine learning is a branch of artificial intelligence. Although it has become a buzzword, it is basically a process of extracting models from data.
Machine learning algorithms use available datasets to develop a model. This model can then make predictions based on new data that was not part of the original data set.
Coming back to our climate problem, there are two main approaches through which machine learning can help us deepen our understanding of climate: observations and modeling.
In recent years, the amount of data available from climate observations and models has grown exponentially. It’s impossible for humans to go through everything. Fortunately, machines can do this for us.
Observations from space
Satellites continuously monitor the ocean’s surface, giving scientists useful insight into changing ocean flows.
NASA’s Surface Water and Ocean Topography (SWOT) satellite mission – slated to launch late next year – aims to observe the ocean surface in unprecedented detail compared to current satellites.
But a satellite cannot observe the whole ocean at once. He can only see the part of the ocean below. And the SWOT satellite will need 21 days to travel to all points of the globe.
Is there a way to fill in the missing data, so that we can have a complete overall picture of the ocean surface at any given time?
This is where machine learning comes in. Machine learning algorithms can use the data retrieved by the SWOT satellite to predict missing data between each SWOT revolution.
Obstacles to climate modeling
Observations tell us about the present. However, to predict the future climate, we must rely on comprehensive climate models.
The latest IPCC climate report was informed by climate projections from various research groups around the world. These researchers ran a multitude of climate models representing different emission scenarios that produced projections hundreds of years into the future.
To model the climate, computers superimpose a computing grid on the oceans, atmosphere and land. Then, starting with today’s climate, they can solve the equations for the movement of fluids and heat in each cell of this grid to model how the climate will change in the future.
The size of each grid cell is what we call the “resolution” of the model. The smaller the box size, the finer the flow details the model can capture.
But running climate models that project hundreds of years brings even the most powerful supercomputers to their knees. So we are currently forced to run these models at coarse resolution. In fact, sometimes it’s so rude that the flow looks nothing like real life.
For example, ocean models used for climate projections generally look like the one on the left below. But in reality, the ocean flow looks a lot more like the picture on the right.
Here you can see the ocean surface currents modeled at two different resolutions. On the left is a model similar to those typically used for climate projections. The model on the right is much more precise and realistic, but is unfortunately too restrictive in terms of calculation to be used for climate projections. COSIMA, Author provided.
Unfortunately, we currently lack the computing power to run high-resolution, realistic climate models for climate projections.
Climatologists are trying to find ways to integrate the effects of the fine, small-scale turbulent movements in the image above on the right into the coarse-resolution climate model on the left.
If we can do that, we can generate climate projections that are more accurate, but still computationally achievable. This is what we call “parametrization” – the holy grail of climate modeling.
Simply, that’s when we can get a model that doesn’t necessarily include all the features of complex, smaller-scale streams (which require huge amounts of processing power) – but can still incorporate their effects in the overall model in a simpler and cheaper way.
A clearer picture
Some parameterizations already exist in coarse resolution models, but often do not do a good job of integrating smaller-scale flow characteristics efficiently.
Machine learning algorithms can use the output of realistic high-resolution climate models (like the one on the right above) to develop much more precise parameterizations.
As our computational capacity increases, along with our climate data, we will be able to use increasingly sophisticated machine learning algorithms to sift through this information and provide improved climate models and projections.
This article first appeared on The Conversation.
Navid Constantinou is an ARC DECRA researcher at the Australian National University.