Machine learning is making inroads into chemical engineering, where it’s helping researchers develop new nanoporous materials for a variety of applications.
Cory Simon, assistant professor of chemical engineering at Oregon State University, is working with Xiaoli Fern, associate professor of computer science, to develop machine learning algorithms to predict the properties of metal-organic frameworks, or MOFs. MOFs are crystalline solids with nano-sized pores. They are constructed from metal ions linked together with organic molecules to form an extended network.
The MOF structure creates a huge internal surface area folded up into a tiny volume. Since gas molecules stick to surfaces, MOFs soak up gas like a sponge soaks up water. By tuning the size and shape of the pore, the MOF can be designed to attract different gases. This lends MOFs to a wide range of applications in gas storage, separation, and sensing.
“MOFs can be used, for example, to increase the density of hydrogen gas, a clean fuel, for storage onboard vehicles,” Simon said. “They can be used to capture carbon dioxide from the flue gas of coal-fired power plants, to prevent it from going into the atmosphere and contributing to climate change. By integrating MOFs into sensors, they can be used to detect dangerous gas leaks. There are all kinds of tasks where nanoporous materials can be quite useful.”
Since many combinations of metals and linkers can be used to make MOFs, many diverse structures are possible and, as yet, unexplored. More than 90,000 have been synthesized, and millions have been predicted. It’s practically impossible, Simon says, to synthesize and test all of them.
“Among a practically infinite number of possible structures, each with their own unique combination of pore shapes and sizes and internal surface chemistries, how can we predict which MOF structure will be optimal for any given application?” he said. “That’s where machine learning comes in.”
One area where machine learning can offer a lot of help is predicting a given property of a MOF on the computer, without expending the often exorbitant resources to actually synthesize and test it in the lab.
“You have materials with known structures, and you can measure the particular property that you care about,” Fern explained. “So, you can collect a bunch of data, pairing materials with this property measurement. Then you can use machine learning to try to learn a mapping from a structure that you haven’t seen before to its property. The trained model can then predict, ‘If I made this material, what is going to be the property associated with it?’”
Simon and Fern are also creating an artificially intelligent “recommendation system” for MOFs. This works much like how Netflix uses AI recommendation systems to predict which new movies a particular user is likely to enjoy.
“Netflix makes movie recommendations for a specific user, based on the data it has about that user’s previous reactions to other movies and their similarity to other users,” Simon said. “Analogously, we are using machine learning to predict missing properties of a material, based on measurements of its other properties and its similarity to other materials.”
Fern and Simon’s future research directions include inverse design and explainability.
“If you have a property that you care about, you want to find the material that optimizes it,” Fern said. “You can start with the existing set of properties of materials you have previously experimented with and measured and use those as kind of a learning basis. You learn from those, and then you try to optimize your material, so that you can actually get the best value for that property. We haven’t actually gotten to that point yet, but that’s on the horizon.”
Fern and Simon are also interested in learning how to distill knowledge from machine learning models to understand the principles that differentiate structures in terms of their properties.
“When you think of AI, it’s often this sort of ‘black box’ concept, where you don’t always know why you got the results that you did,” Fern said. “We want to see if machine learning can help us understand the principles that differentiate different structures. So, another, deeper question is how can we understand and interpret the model to gain knowledge?”
Explainability confers greater credibility on the predictions generated by machine learning, ultimately making them more useful to researchers.
“If we tell a chemist, ‘The machine learning algorithm said you should invest three months of your time on this material,’ they will want to know: Why is the machine learning model thinking that?” Simon said. “Explainability, as well as uncertainty quantification, can help motivate chemists to follow up on our predictions.”