(129c) Machine Learning-Aided Design of Biodegradable Polymers

In the pursuit of sustainable materials, biodegradable polymers have emerged as promising alternatives to traditional plastics, finding applications across diverse industries. Particularly, the degradation products of traditional plastic are of specific concern, and fully biodegradable and non-toxic plastic alternatives offer promising solutions to this ongoing challenge. However, a rational design approach to engineer biopolymers for degradation after their intended use still remains elusive. This is largely due to our inability to understand and model performance metrics (such as weight loss over time) capturing degradation behavior of these materials as a function of chemical, geometric and environmental factors.

As an exciting development in this direction, we have established predictive machine learning models which utilize physics-informed deep neural networks (NN) on previously established, manually curated experimental data that characterizes the weight loss behavior of fully biodegradable polyester copolymer samples in both water and soil natural environments. The models were applied in a series of experiments to predict the mass loss of approximately 10,000 structural and compositional variations of the 230 homo- and copolymers originally included in the original training set. The predicted mass loss for these candidate biodegradable polymers over a period of 365 days allowed us to identify novel copolymers with the potential to replace existing chemistries while matching property values to the desired performance metric ranges. This talk will discuss our findings and future directions, including the integration of critical properties, such as, thermal and mechanical characteristics, into the screening and design workflow and thereby bringing us one step closer to realizing the grand vision of a sustainable circular plastic economy.