Symposium X—MRS/The Kavli Foundation Frontiers of Materials

Juan J. de Pablo, New York University

AI-Enabled Design of Sustainable Polymeric Materials

Written by Molly McDonough

Plastics have been being sold, used, and researched since their advent during the Industrial Revolution in the 1800s. Now, we find plastics everywhere, from our clothes to airplanes’ coatings. In 2024, over 220 million tons of plastic waste will be generated, and as many of us know, plastic pollution is a large concern. Plastic pollution can find its way into drinking water, food, soil, and our bodies, threatening the health of humans, other animals, and the environment. Juan J. de Pablo’s research into the artificial intelligence (AI) design of polymetric materials aims to find more sustainable polymer materials to help mitigate the issues caused by plastic pollution. However, the major constraint on this development is the already existing industrial processes corporations worldwide have implemented to manufacture plastic. It would be too expensive and time-consuming to deploy new manufacturing methods for sustainable polymers, so de Pablo seeks to find new polymetric materials that can fit into existing industrial processes. Additionally, de Pablo aims to create new materials with the same or better properties, costing the same or less than conventional plastics like polylactic acid (PLA), polyethylene, and polyester.

De Pablo’s group looks for sustainable polymer candidates by developing tools to predict the rheology of polymers and to design polymers with particular rheological responses. This approach begins by designing new monomer molecules with targeted equilibrium processes. The first step in this process was asking if scientists could use existing large language models (LLMs) such as ChatGPT and retrain them using existing polymer databases such as simplified molecular-input line-entry systems (SMILES) to predict properties of molecules. Unfortunately, current-generation LLMs lack the domain-specific fine-tuning for this application, impacting their ability to understand and predict chemical language accurately. At best, ChatGPT could predict chemical language with a 20% success rate, which is not high enough to be reliable to conduct experiments based on it.

Instead, de Pablo implemented a different approach, using variational autoencoders (VAE). VAE combines known chemistry in the form of functional groups with graph neural networks (GNNs). Then, the GNNs are applied to encode information at two levels, atom graph and motif graph, to encode the structure of the polymer fully. Next, they implement multilayer perception for hierarchical connections between the graphs. After this, the decoder operates in an autoregressive manner. Finally, the model reconstructs the molecule by iteratively generating a new motif and connecting it to the partial molecule built thus far. This process resulted in models that are chemically meaningful; despite using unlabeled data (unsupervised learning), the model is able to cluster molecules based on their chemistry. This allowed de Pablo to simultaneously model the density, cohesive energy, radius of gyration, heat of vaporization, glass transition temperature, and isothermal compressibility, which are all important parameters in creating new, usable, polymetric materials while using a small dataset of only 600 materials. There are still two challenges with this approach: (1) being that it only simulates micromonomers, meaning that there are millions of ways to arrange them that yield different properties, and (2) the model can predict materials that are not thermodynamically stable or able to be synthesized.

As such, de Pablo went back to the drawing board and decided to implement this machine learning framework in conjunction with a dataset of polymers. Here, his research team uses a database of standard polymer structures, each with primitive units of 4 or fewer elements, and N = 400 along the backbone. They find here that after about 500 iterations through polymer structures, the model is highly optimized, likely due to the dependence on property and chemical complexity. This machine learning framework can then be easily repurposed for design, and Bayesian optimization yields numerous successful candidates in a space not included in the training data.

The last portion of de Pablo’s work focuses on creating a database of dimers, trimers, some oligomers, and probabilistic representations, combined with experimental data to create a multiscale approach to predicting rheology and processing techniques of polymer materials. This database aims to integrate all atom scale, coarse-grained, and slip spring simulations to fit experimental data and yield new potential polymer materials. De Pablo closed his talk by discussing using this framework to design degradable polymers. The model takes a simple polyethylene and intercalate monomers susceptible to degradation. This allows de Pablo to determine how degradable monomers behave in existing polymer structures, allowing us to understand if the monomers remain degradable when integrated into a larger system. This work extends existing machine learning frameworks to discover new materials that have the potential to solve critical issues in plastics manufacturing and recycling.

Symposium X—MRS/The Kavli Foundation Frontiers of Materials features lectures aimed at a broad audience to provide meeting attendees with an overview of leading-edge topics.