Putting hydrogen to work

Scientists, industries and policymakers have been working for years to find energy storage technologies that will help meet peak demand for transportation, power generation and other energy-hungry activities.

Take, for instance, hydrogen, a promising option for energy storage that is stable, portable and pollution free. Most of the current hydrogen supply is generated by splitting methane, which comes from natural gas, in a process called pyrolysis. Because of increasing energy demands, researchers and industry leaders are looking for methane alternatives.

Alternative chemical generation of bulk hydrogen, however, carries big challenges. Converting compounds to hydrogen without injecting energy would be “completely unfeasible because the molecules don’t want to do anything — unless you make it easier for them to react by reducing the energy barrier,” explains Boris Kozinsky, a computational researcher at Harvard University. The answer: designer catalysts and optimized microscopic environments that speed up molecular interactions and chemical reactions.

Improving industrial-scale catalysis hinges on atomic-scale understanding as you tinker with potential improvements, says Albert Musaelian, a researcher in Kozinsky’s group and a recent graduate of the DOE Computational Science Graduate Fellowship. Moving from natural gas derived methane as a hydrogen source, he says, means finding new chemical inputs — whether it be captured from the air, or from chemically diverse renewable sources such as landfill gas or biomass.

At Harvard, Kozinsky’s group is applying machine learning and supercomputing to study new ways to catalyze chemical reactions that produce hydrogen. Kozinsky’s interest in complex chemical systems was borne from a decade-long career in industry before he transitioned to academia. “Everything is made of atoms, and all we’re doing is calculating how atoms move around,” he says. “We’ve been working on developing new types of approaches to molecular dynamics and scaling them to simulate large and complex systems for a long time, so we can get as close to reality as possible.”

This year, Kozinsky and his group were awarded a grant from the Innovative and Novel Computational Impact on Theory and Experiment (INCITE) program to use up to 150,000 node hours on the Argonne Leadership Computing Facility’s Polaris and the Oak Ridge Leadership Computing Facility’s Frontier exascale computing system.

Kozinsky’s team focuses on two main reactions: dehydrogenation, in which hydrogen is a chemical breakdown byproduct; and hydrogenation, in which it’s added to molecules. They want to model how molecules move and rearrange in real time, a simulation method called molecular dynamics. To do that, they need to know the physical forces on the atoms. In principle, they could determine those forces using quantum mechanics, but such computations are sluggish and expensive. Their machine-learning algorithms are orders of magnitude faster and trained on the results of quantum mechanical computation.

‘Unlike experiments, we can know everything that’s happening to molecules on our surface.’

One challenge that Kozinsky’s team has faced is implementing these algorithms to work specifically on different supercomputers, as many of them are based on GPUs from various hardware vendors.

So far, Kozinsky and his colleagues have simulated catalysis reactions on supercomputers, reaching up to a trillion atoms in scale, and in some cases over nanoseconds. “That sounds very short,” explains former graduate student Anders Johansson. “But the typical vibrational scale of a hydrogen molecule is femtoseconds,” or a millionth of a nanosecond.

“For the first time, we were able to run a direct simulation of a catalytic process that hasn’t really been done before to the extent that it actually agreed with experiments,” Kozinsky says. “So when we run a simulation of hydrogen gas reacting on a platinum-catalyst surface, we can count all the reactions that happened in a simulation with a billion or a trillion atoms.”

Within this framework, he and his colleagues can now vary temperature, surface dimensions and other key features, then see how these results might square with experimental data. (Their collaborators, such as Anatoly Frenkel at Brookhaven National Laboratory, are validating their results using experimental methods such as X-ray spectroscopy.)

Kozinsky finds this early agreement with experimental data promising because it suggests that “you can do a thousand different calculations with different catalyst structures, temperatures or pressures and test out many different conditions. At the same time, unlike experiments, we can know everything that’s happening to molecules on our surface.”

Some of their early results from the INCITE grant have also disproven a common assumption that catalyst surfaces are ideal and static (as summarized in a preprint earlier this year). “It’s a very complex, coupled dynamic system where the molecules and the catalyst change in response to each other,” Kozinsky says. He hopes his research can better inform the design of catalysts to produce hydrogen on an industrial scale in the future.

“We’re very quickly moving to a point where the design of new catalysis can be accelerated using these combined and experimental methods,” Kozinsky says. “We’ve demonstrated with our codes that they can run on supercomputers so we can answer real technological questions.”

Under the same INCITE grant as the catalysis work and using the same machine-learning methods, Kozinsky is studying the chemical reactions in batteries. He and his colleagues hope that this research can help companies develop and design souped-up batteries that resist degradation and offer a way to decrease costs to power electric devices.

Lithium-ion batteries, for instance, are a necessity for many electronics, and researchers are working to improve charging speed and increase battery life. But the rate at which such batteries currently degrade, and the risk of catastrophic consequences including fires, is holding back innovation. Currently, the electrolyte in the batteries is an inflammable organic liquid, and other researchers are looking into replacing that with non-burning inorganic oxides or sulfides. Simulation may be the key to building safe, high-performance batteries for an array of applications.

“The scope of possibilities right now is really wide,” Kozinsky says. “We’re extremely excited by the possibility that we’re opening with these methods of designing materials.”