Generative AI for polymer design passes lab tests with a new dielectric material

Schematic workflow illustrating the LLM-based framework for organic material design targeting dielectric applications. Credit: npj Artificial Intelligence (2026). DOI: 10.1038/s44387-026-00087-1

The words on this page mean something because they are assembled in a particular order and follow the complex rules of grammar and syntax. Creating new chemical polymers follows a similar kind of structure, with rules about what elements and groups of atoms go together and how to assemble them to make sense. Thinking about polymers in that way has led Georgia Tech materials scientists to create new generative artificial intelligence tools that are like Claude or ChatGPT for new materials.

These are the first foundational models for generative polymer design that have also been validated through physical experiments: users specify the properties they need in a polymer and the model will suggest a chemical structure.

Led by Regents’ Entrepreneur Rampi Ramprasad, the researchers published their latest model this month innpj Artificial Intelligence—including a test material they created and validated in the lab to prove that the model works.

“This architecture learns the chemical semantics and chemical grammar. It learns what is allowed, what is not allowed, what comes together well, and what makes a good chemical sentence,” said Ramprasad, Michael E. Tennenbaum Family Chair and professor in the School of Materials Science and Engineering. “A word is really defined by the neighbors it keeps. That’s the context for the meaning of a word. It’s the same with atoms or clusters of atoms.”

The new Georgia Tech tool, called POLYT5, overcomes barriers that have limited previous generative AI approaches to polymer design: they sometimes suggested polymers that failed to follow the rules of chemistry or couldn’t be created in real-world labs.

“The chemical structures POLYT5 generates are 100% robust and will follow chemical grammar and semantics,” Ramprasad said. And because the model is trained only on polymers that can realistically be made, its ideas are more likely to be usable.

To create POLYT5 and a related tool called polyBART, Ramprasad, Research Scientist Harikrishna Sahu, postdoctoral scholar Wei Xiong, and Ph.D. students Anagha Savit and Shivank Shukla started with existing AI language architectures. They stripped out the natural language training and instead taught the model using polymer chemical structures—the words and sentences of chemical language. They used more than 12,000 experimentally produced polymers from research studies, plus a bank of more than 100 million hypothetical candidates.

To test the new models, the research team asked them to suggest designs for materials called polymer dielectrics, which are used in electric vehicles, defibrillators, and other places requiring a quick burst of energy. The team outlined a series of specific properties that would result in good performance at high temperatures and allow the material to be processed industrially.

The researchers then picked one of the top candidates from each model’s output and created it in the lab. Their testing showed they performed just as expected.

Though Ramprasad and his team fine-tuned the base model for dielectrics in the study, the idea is that it can be tuned for any combination of properties. In fact, his group is doing exactly that, teaching the model to design polymers for very different uses as part of new projects underway in his lab.

Discover the latest in science, tech, and space with over 100,000 subscribers who rely on Phys.org for daily insights.
Sign up for our free newsletter and get updates on breakthroughs,
innovations, and research that matter—daily or weekly.

In the npj Artificial Intelligence study, the researchers also paired POLYT5 with a general-purpose large-language model. The goal is to make it possible for more scientists to use it without being experts in polymer design and creation. “We live in a world where people want to be able to converse with these agents, and we ultimately want to decrease the barrier for adoption of these tools,” Ramprasad said.

POLYT5 and polyBART build on years of previous work in Ramprasad’s lab. His team has built tools to predict properties of potential polymers to speed discovery of new materials. The new model offers the inverse: using the desired properties to design the polymer from the beginning.

“Language, they say, is the frontier of intelligence. And this is a different kind of language—the language these atoms follow,” Ramprasad said. “I’ve worked in this field for a long time, but even five years ago, I would not have thought that language models could be adapted to learn chemistry.”

More information

Harikrishna Sahu et al, POLYT5: an encoder-decoder foundation chemical language model for generative polymer design, npj Artificial Intelligence (2026). DOI: 10.1038/s44387-026-00087-1

Provided by
Georgia Institute of Technology

Citation:
Generative AI for polymer design passes lab tests with a new dielectric material (2026, March 26)
retrieved 26 March 2026
from https://phys.org/news/2026-03-generative-ai-polymer-lab-dielectric.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.