Design of (De)Polymerizable Polymers Using Machine Learning-Based Predictive Models and Generative Algorithms
Author(s)
Kern, Joseph Daniel
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Plastics stand as one of the most ubiquitous materials in modern society, with production exceeding a staggering 400 million metric tons in 2022 alone \cite{GlobalPlasticProduction}. Regrettably, the inherent thermodynamic stability of the current generation of plastics poses a significant challenge, rendering many incapable of effective recycling. Consequently, these plastics persist as waste, potentially lingering for hundreds of years. Thus, there arises an urgent need for the development of novel plastics that not only satisfy the demands of diverse applications but also possess the crucial capability to be readily recycled. It is within this context that the primary objective of this research is situated: to identify promising candidates capable of replacing contemporary commodity plastics with chemically recyclable (depolymerizable) alternatives. To achieve this overarching goal, my work focused on several critical components:
\begin{enumerate}
\item\textbf{The development of an AI-enabled Virtual Forward Synthesis (VFS) platform aimed at generating novel plastics from existing molecules}: This platform enables automated searches for molecules capable of facilitating ring-opening polymerization (ROP), as ROP polymers are recognized as promising candidates for depolymerizable designs due to their unique thermodynamics. Leveraging the power of machine learning (ML), the platform further predicts polymer properties and identifies promising candidate molecules, thus paving the way for the discovery of recyclable plastic alternatives.
\item \textbf{The creation of a genetic algorithm to rapidly explore polymer design spaces}: This algorithm was crafted to specifically cater to ROP chemistries, marking a significant advancement over previous versions. It is capable of swiftly identifying promising candidates from innumerable search spaces, accomplishing this task with remarkable efficiency compared to enumerative design approaches.
\item \textbf{Advancement of best-in-class ML models for predicting solubility and toxicity}: These factors wield substantial influence over polymer processing and the environmental toxicity associated with them. By harnessing the predictive capabilities of these models, promising polymer candidates can be further refined, thus honing in on better polymer designs.
\end{enumerate}
Employing these developed methods, a myriad of candidate polymers emerged as promising alternatives to one of the most ubiquitous commodity plastics, polystyrene (PS). Among these candidates, one is currently undergoing synthesis exploration by polymer chemists. Furthermore, invaluable insights were gleaned on the development of thermally and mechanically resilient polymers. Additionally, a plethora of software packages and models have been made readily accessible for use.
Sponsor
Date
2024-07-01
Extent
Resource Type
Text
Resource Subtype
Dissertation