Princeton researchers have developed a new tool to speed the discovery of advanced materials known as metal organic frameworks, or MOFs.
MOFs are an emerging class of materials that form microscopic sponge-like structures with vast interior surface area. That quality promises to transform how society traps, absorbs and filters substances at the molecular level. The researchers say this could lead to better battery chemistry, more efficient carbon capture and improved access to clean water.
But scientists face a problem of choice. MOFs are highly modular, consisting of metal-ion nodes and organic molecules that link the nodes into large networks. The researchers say there are trillions of possible chemical combinations. But not all combinations are equally useful, and some are not even feasible to make in a lab.
Now, a team led by Adji Bousso Dieng has developed a method using machine learning to predict which MOF structures are good candidates, avoiding the need for researchers to wade through countless useless structures.
"Our tool takes seconds to produce a prediction versus between seven hours to more than two days for traditional molecular simulations," said Dieng, assistant professor of computer science, and associated faculty member in the Princeton Materials Institute.
Specifically, the team used a measure that describes the stability of a molecular structure — called free energy — to make the predictions.
They published a paper detailing the work on December 16 in the Journal of the American Chemical Society.
Developing the tool involved three major steps. First, they had to convert key physical and chemical characteristics of MOFs into sequences that a machine could read. Second, they assembled a database of MOFs on which to train their model. Finally, they ran their model multiple times to make predictions about specific materials.
Turning physical and chemical characteristics related to the free energy of the individual atoms and units of a MOF into a machine-readable representation was arduous, and key to this research.
"The sequence representation we came up with is really what unlocked everything," said Dieng.
Using this system, they generated representations for one million MOFs.
The team then trained a bespoke language model to predict the free energy values of all one million MOFs in their database, calibrated using a simpler characteristic that is closely related to free energy. When they tested a sample of the database, using around 65,000 materials with known free energy values, the model's predictions were accurate 97 percent of the time.
Dieng's collaborator, Diego Gómez-Gualdrón from the Colorado School of Mines, had previously determined that, below a certain free energy value (4.4 kilojoules per mole), a MOF is considered stable and can be feasibly synthesized in a lab.
"If you have a new MOF, you can predict its free energy, and you can also predict whether it is synthesizable or not," said Dieng.
The team is now working on streamlining their sequence representations, reducing the computational overhead incurred with some structures. They are also using that system to add a search function to their tool to help find stable MOFs.
"We are lifting the problem where now you can compute the sequence representation itself very quickly, very cheaply," said Dieng. "This technology allows researchers to focus resources on promising candidates for practical applications in carbon capture, energy storage, catalysis and gas separation."
The article, Highly accurate and fast prediction of MOF free energy via machine learning , was published on December 16 in the Journal of the American Chemical Society. Besides Dieng and Gómez-Gualdrón, authors include Andre Niyongabo Rubungo from Princeton University and Fernando Fajardo-Rojas from the Colorado School of Mines. Support was provided by the National Science Foundation and the Schmidt Sciences AI2050 Fellowship.
