Facebook deploys its AI to find green energy storage solutions

'We're hoping to basically replace DFT with a machine learning algorithm.'

koto_feja via Getty Images

Our traditional solution to the unpredictable nature of renewable energy sources like solar and wind power has generally been to simply dump the excess wattage back into the local grid or sequester it away in utility-scale batteries. But as more and more of our power generation is created by renewables, their production capacities can potentially outstrip that of the local grid while battery technology can quickly become prohibitively expensive at scale.

One alternative is putting that excess power to work driving catalytic reactions. “There are a lot of different ways that we can store the energy,” Zack Ulissi, CMU Assistant Professor of Chemical Engineering and Materials Science and Engineering, told Engadget. “The most well known is you take water and you electrolyze it to split it into hydrogen and oxygen. And then you can take that hydrogen and run it into a hydrogen fuel cell.” We could also use it to produce methanol, a basic feedstock for the chemical industry, or make ethanol for use as a liquid fuel, he added. We could even re-catalyze the ethanol, directly stripping the electrons from the respective molecules with any combustion, thanks to a 2019 discovery out of the University of Arkansas and Brookhaven National Laboratory.

But in order for the catalytic process to be viable and effective, the catalyst has to be as efficient as possible. However, given that catalysts are typically made from a 3- or 4-element combination out of a field of nearly 50 potential elements and, when taken together with a litany of other chemical and structural variables -- from the ratio of ingredients to the configuration of the elements to the catalyst’s physical surface shape -- there are billions of potential ways to build a “best” catalyst. And that’s just for a single chemical reaction.

The process for investigating new potential compounds is therefore quite slow, so to help accelerate the catalyst creation process, Facebook AI has teamed with Carnegie Mellon University on the Open Catalyst Project. They plan to train machine learning algorithms on open-source data to “accurately predict atomic interactions dramatically faster than the compute-heavy simulations scientists rely on today,” according to a Wednesday blog post by the company.

Those simulations include Density Functional Theory (DFT) a quantum mechanical system chemical engineers will often employ to find the most promising candidates and avoid potential research dead ends.

“Density Functional Theory is one way of solving for how electrons interact in the system, you basically try to calculate the electron density,” Ulissi explained. “DFT says that, if you can get the electron density, then you actually know the final energy [of the system].”

Now, if you think that using quantum mechanics to simulate the relative movements of electrons, atoms and molecules in a many-body system in an attempt to find the configuration with the lowest final energy (the “relaxed state”) requires an absurd amount of processing power and time to compute, you are very much correct. Even with access to Facebook’s high-end servers, the relaxation calculations can take anywhere from 12 to 72 hours for a single candidate material iteration.

So rather than try to brute force these billions of permutations, Facebook and CMU have established the Open Catalyst 2020 data set, a collection of 1.3 million relaxations of molecular adsorptions onto surfaces and plan to leverage it to train machine learning algorithms on the “fundamental physics governing quantum mechanics, teaching the models to approximate the energy and forces of molecules based on past data.”

Not only should such a large data set help dramatically improve machine learning models’ generalization capabilities but also teach them “the underlying physics governing molecules at inorganic interfaces,” Larry Zitnick, Facebook AI research scientist argued in the Wednesday blog post.

If the researchers are successful in training an ML model, Zitnick noted to Engadget, “we can take something that used to take eight hours and it's likely that we may be able to make those same relaxations and do them in under a second… We're hoping to basically replace DFT with a machine learning algorithm.”

“Really what we're thinking about is how can we get AI or ML to help us in this process of selecting the [candidates] that are most interesting,” Ulissi said, “where we spend our time, or helping us identify trends and get on higher level insights into what's going on so that we can guide the experimentalist in one direction.”

“That's what we're aiming for,” Zitnick replied. “We want to personally start doing large scale exploration of catalysis, not just test it 10s of, or 10,000, or 100,000, or millions, but start testing billions of different possibilities.”

As such, the Open Catalyst 2020 data set has been made open source and available to the research community. Zitnick hopes to put together a Facebook Challenge using the data set in the near future.