Facebook A.I. researchers push for a breakthrough in renewable energy storage

October 14, 2020, 1:00 PM UTC

Facebook is one of the world’s largest corporate purchasers of renewable energy, having signed deals for more than 5 gigawatts of “clean” power—3.5 gigawatts of which is expected to be online by the end of the year. (One gigawatt is enough to power 700,000 homes for a day, according to some estimates.)

The company uses all that energy to power the vast data centers that run its social media platforms, as well as to keep its various offices humming—though Facebook, along with other large technology firms, has been faulted by environmental groups for the sheer amount of energy its data centers consume.

Now Facebook seeks to solve one of the biggest impediments to making power grids more green: the need to balance the irregular supply of power from renewable sources like wind turbines and solar panels more evenly so that it can match demand.

A major drawback of most clean energy sources is that they are only available intermittently—solar power when the sun shines and wind power when it’s windy. Often those conditions don’t line up with demand. As a result, most power grids still have to rely on fossil fuel–based power generation or nuclear power to ensure the grid has enough power, no matter the weather.

Storing excess renewable energy in batteries for use later is one potential solution to this problem. But even with a breakthrough in battery technology, it is likely to be prohibitively expensive and unwieldy.

Another option is to convert renewable electricity into some other form of relatively clean energy, such as hydrogen, ethanol, or ammonia. But achieving this today involves chemical processes dependent on expensive catalysts, such as platinum, and the processes remain fairly inefficient, which means that a good portion of the electric energy goes to waste.

Researchers at Facebook think artificial intelligence might be useful for finding better catalysts. They’ve teamed up with chemical engineers from Carnegie Mellon University in Pittsburgh on an initiative they hope will drive innovation in the field.

On Wednesday the group released a data set of 1.3 million simulations of the quantum mechanical interactions between molecules of two chemical compounds when they come into contact. Called Open Catalyst 2020, it is the largest such data set ever assembled and should enable researchers to train a machine learning algorithm to begin to make accurate predictions about these chemical interactions, according to a Facebook blog post.

Zachary Ulissi, a chemical engineering professor at Carnegie Mellon who is working on the project with Facebook, said that simulating the atomic-level interactions of two chemicals, particularly those with higher atomic masses, is extremely complicated and time-consuming using existing methods, which mostly involve a technique called density functional theory (DFT). Each simulation takes weeks to months, he said. “It’s very hard because of the number of electrons in the system.”

Modern chemistry labs, using very powerful computers, can use DFT to simulate at most 40,000 compounds per year, Facebook said in its blog. But there are billions of possible combinations to explore to find candidates that might make effective catalysts—and even after finding one that looks promising in simulation, it must still be validated through real experimental chemistry in the lab.

Facebook said that using the existing DFT process running on the company’s own high-performance computers, it took about 12 to 72 hours to create each set of simulations that is in the Open Catalyst data set.

The company said that by using a machine learning method based on deep neural networks—a kind of A.I. software loosely modeled on the human brain— it hopes to be able to train an algorithm that might be able to make accurate predictions about the catalytic potential of two compounds in just seconds.

Facebook is also making the Open Catalyst data set freely available to other researchers to use to train A.I. algorithms, and it is creating a leaderboard to rank how well various algorithms perform at predicting known catalytic interactions. In other areas—from predicting how proteins fold to figuring out how to detect hate speech—this kind of competition for bragging rights has proved to be a powerful incentive that has spurred progress toward creating powerful A.I. algorithms that can tackle real world problems.

Larry Zitnick, a Facebook A.I. researcher based in Menlo Park, Calif., who is helping with the catalyst project, said the company’s scientists thought they could make an important contribution to the search for better catalysts for two reasons: one is the sheer amount of computing power that Facebook has available. The other is the Facebook A.I. research lab’s expertise in a particular kind of neural network called a graph neural network that is good at capturing complex relationships and interdependencies between variables. Zitnick said he thinks graph neural networks may be especially useful in predicting catalytic interactions, although this has yet to be proven.

Ulissi said that chemical engineering had only recently begun to figure out how to apply A.I. techniques to some of its biggest problems. He pointed to recent progress that has been made in training algorithms to accurately predict the shape and structure of small molecules that might be useful in creating new medicines as an example of what might now start happening with chemical catalysts too.

But he cautioned that it is even more difficult to make accurate predictions about the subatomic interactions of the large inorganic compounds and metals useful for converting electricity into energy stored in another type of fuel than it is to make forecasts about the small organic molecules often used in pharmaceuticals.

This post has been updated to clarify that the 12 to 72 hours mentioned in the story refers to the time it took Facebook to simulate each catalyst included in the Open Catalyst data set using the DFT method, not the training time for a neural network to learn to accurately predict a catalytic interaction.

More must-read tech coverage from Fortune: