Machine learning and artificial intelligence have accelerated the ability to design materials with specific properties. But although scientists have succeeded in designing new metal alloys, polymers – such as plastic used for bags – have been much more difficult to design.
Researchers at the Pritzker School of Molecular Engineering (PME) at the University of Chicago have found a solution for polymer design by combining modeling and machine learning.
Through the computerized construction of nearly 2,000 hypothetical polymers, they were able to create a database large enough to train a neural network – a type of machine learning – to understand what polymeric properties arise from different molecular sequences.
“We show that the problem can be solved,” said Juan de Pablo, a Liew molecular engineering professor who led the research. “Now that we have established this basis and shown that it can be achieved, we can take steps forward in using this framework to design polymers with specific properties.”
Creating a database for learning polymer sequences
To create the database, the researchers used almost 2,000 polymers constructed on a computational basis, all with different sequences and performed molecular simulations to predict their properties and behavior.
When they first used a neural network to find out which properties are based on which molecular sequences, they were not sure if they would find a reasonable answer.
“I didn’t know how many different polymer sequences are needed to learn the behavior of materials,” de Pablo said. “The answer could have reached millions.”
Fortunately, the network needed only less than a few hundred different sequences to find out the properties and predict the behavior of completely new molecular sequences. This meant that experimentalists could now follow a similar strategy and create a database to train a machine learning network to predict the properties of polymers based on experimental data.
However, that was only half the problem. Subsequently, the researchers had to use the information learned by the neural network to design new molecules.
They continued this and, for the first time, were able to demonstrate the ability to choose a property from a polymer molecule and use machine learning to generate a set of sequences that would lead to those properties.
Design of specific polymers
Although the system has been trained to understand only one type of polymer, the potential implications could be extended to several types. Not only could large companies design greener products, but they could also design polymers with the properties they want.
Polymers are usually dissolved in solvents for paints, cosmetics, medicines, medical solutions and foods to control the flow of liquids, for example.
Polymers are also used in a wide range of advanced technologies, from aerospace applications to energy storage, to electronic and biomedical devices. Designing high-precision polymers for specific applications could allow companies to design materials in a more accessible, easier, and sustainable way.
In the future, the research group hopes to involve experimentalists in the development of some of the polymers they have designed and to continue to refine their system to create even more complex polymers. Relying on robotic systems for high-speed synthesis and characterization of new molecules, they hope to expand their database to include experimental data.