(149a) High-Throughput Measurement of a Machine Learning Model for Polyester Biodegradation | AIChE

(149a) High-Throughput Measurement of a Machine Learning Model for Polyester Biodegradation


Av-Ron, S. - Presenter, Massachusetts Institue of Technology
Olsen, B., Massachusetts Institute of Technology
The U.S. generates more than 35 thousand tons of plastic waste every year and less than 20% is recycled, leading to an excessive amount of plastic accumulating in landfills. Many used plastics never make it to controlled end-of-life locations and are left in the environment to degrade, which can take more than a century. There is a dire need for degradable alternatives that can compete with current everlasting plastics. Currently, the study of compostable and biodegradable candidates is a low-throughput process, in part due to the present methods and standards to study biodegradation which require testing times of multiple months and capital-intense measurement equipment. These methods substantially reduce the size of data sets on polymer biodegradation, limiting our ability to develop group contribution theories and data-driven models to predict biodegradation performance.

Here, we combine high throughput synthesis, high throughput biodegradation assays, and machine learning to produce a predictive model for polyester biodegradability. This study utilized the clear zone test to obtain biodegradation rates of hundreds of polyesters using a sample of different bacterial strains. For the clear zone test, polymer is homogeneously dispersed in media, and its degradation during bacterial growth is monitored optically. This biodegradation test was applied to a library of hundreds of polyesters representing diverse synthetic routes and monomer functionalities, including biobased and synthetic systems to generate a large data set for machine learning. Different representations for polymer chemical structure and machine learning algorithms were explored in order to identify key chemical features that promote degradation and to identify models with the greatest predictive capacity.