Designing new polymers with AI — Q&A Overview
By: Arash Khajeh
The TRI Accelerated Materials Design and Discovery team has released two papers on the use of generative AI techniques to design new polymers (here and here).
Our goal is to have polymers used as solid electrolytes in lithium-ion batteries. For decades, it has been recognized that a polymer electrolyte – rather than a liquid electrolyte – could be a game changer for lithium-ion batteries by enabling high energy density with excellent safety. Polymers have the advantage of being mechanically robust and flexible while being resilient to extreme temperatures and pressures. The challenge is that the lithium conductivity in polymers is too low to be practical in batteries. The best-performing dry polymers are based on variants of (poly)ethylene oxide, which were discovered almost 40 years ago. While the conductivity has improved, a breakthrough performance has yet to be seen.
TRI’s Senior Director of Energy & Materials, Brian Storey, speaks with Arash Khajeh, a Research Scientist at TRI, about the team’s work and the future prospects of using generative AI to create new polymers that might break the conductivity barrier.
How are you using advances in generative AI, such as GPT-based models, to make hypothetical polymers?
We start with a text-based representation of polymers that includes information about the composition and connectivity of atoms in the monomer. Just as with written language, these models can learn the “language” and “syntax” of molecules. The model can propose new alternative chemical structures for further exploration. To make the generative models useful, we also need to incorporate the target properties of polymers into the model input. This type of conditional generation allows us to use a specific class of materials’ performance and force the model to generate better-performing materials, or high ion-conductive polymers in this case.
GPT models are known for major inaccuracies sometimes. How do you prevent the method from generating polymers which are physically infeasible and/or irrelevant?
To avoid generating unrealistic materials, we check for chemical validity and synthetic accessibility of generated polymers. In addition, we continuously compare the generated structures against existing data to identify novel and unique polymers. Although some of these tests, such as synthesizability, were originally developed for small molecules, they can still provide an estimate of the likelihood of synthesis of these newly generated polymers in the laboratories. However, many of the polymers are still not useful due to difficulties in synthesis or instability.
For future development, we are trying to incorporate expert human feedback as a part of the process to improve the usefulness of generated polymer structures. This integration of human expertise and computational power holds great potential for accelerating the discovery of new polymers with desirable characteristics.
You are using molecular dynamics simulation to test the polymers for better conductivity. How do you force the generative model to create better-performing polymers?
Molecular dynamics (MD) simulations allow us to calculate, within reason, whether a new polymer might be a good ion conductor. We use the feedback from MD simulations performed on the generated polymers to strategically select and add promising materials to the training data. The model gradually sees better-performing materials and is biased toward generating similar structures. The closed-loop polymer discovery framework we developed continues to generate better polymers. The generative models trained on MD simulation data can generate polymers that exhibit ionic conductivity higher than PEO based on MD simulations. Even though many of these polymers could be hypothetical, we believe we can learn from ion transport mechanisms in these new polymer systems and use this information to improve the design of real-world polymers.
What are the limitations of this method?
First of all, the trained generative models are bounded by the range of observed polymer structures and their behavior. They are not creative enough to generate completely different structures with new ion transport mechanisms, which is what we might ultimately need for solid polymer electrolytes. Second, the molecular dynamics simulations are optimized for specific polymer structures and are not completely generalizable to a wide range of polymer chemistries and structures. MD simulations have inherent uncertainty, which adds a layer of complexity.
Despite our best efforts, models can only approximate real-world behavior, and there will always be a gap between simulation and experimental results. Continuous refinement of MD simulations for a variety of polymer systems coupled with experimental validation can enhance the versatility and accuracy of our models in the future.
What can this method enable in the future and what are you most excited about?
An exciting aspect of the framework we have developed in these two studies is that we can use it for various classes of materials, properties, and different applications. I think the future research direction in the integration of human expertise and computational power holds great potential for accelerating the discovery of new materials with desirable characteristics, ultimately contributing to advancements in various fields such as materials science, biomedical engineering, and environmental sustainability.
