Researchers at Georgia Tech have developed a machine-learning model that makes it a lot easier for materials scientists to develop new polymers.
Polymers are large molecules made from long strings of chemical building blocks. We use synthetic polymers like Nylon, Polyester, or plastic packaging every day.
Materials scientists try to come up with new arrangements of polymers to make materials that have specific properties. If you wanted a material that was lightweight, stretchy, waterproof, and heat resistant you could make a polymer that had all those properties.
The problem is that figuring out which combination of chemicals would create a polymer with those properties is a huge undertaking. There are endless combinations and it’s extremely difficult to predict what properties a new polymer will have.
The Georgia Tech researchers trained their LLM, called polyBERT, on a dataset of 80 million polymer chemical structures. The result is a model that understands the language of chemicals.
In the same way that LLMs are trained in a language like English, polyBERT now understands the grammar and syntax of how chemicals and atoms combine to create polymers.
The National Science Foundation (NSF) funded the research behind polyBERT. Its program director, Debora Rodrigues, said the researchers were “developing a new artificial intelligence tool to overcome the challenge of determining which combinations of chemicals will make the most effective polymers.”
By using polyBERT, materials scientists can work through combinations of chemicals more than 100 times faster than before. The result of using the model is a dataset comprising 100 million hypothetical polymers and their predictions for 29 properties.
If you’re a materials scientist looking for a new material with very specific properties you don’t need to experiment and hope for the best, you can just consult the dataset that the AI generated.
While the researchers trained polyBERT on polymers, they said it could theoretically be used for other chemical research too.
The potential for creating materials that are more sustainable or energy-efficient presents an immediate real-world benefit from applying this kind of artificial intelligence.
