Machine Learning for Melting Temperature Predictions and Design in Polyhydroxyalkanoate-Based Biopolymers

Diminishing fossil fuel-based resources and ever-growing environmental concerns related to plastic pollution demand for the development of sustainable and biodegradable polymeric material alternatives. Polyhydroxyalkanoates (PHAs) represent an eco-friendly and economically viable class of polymers with a wide range of applications. However, the chemical diversity combined with tunable physical properties available within PHAs poses discovery and optimization challenges with respect to identifying optimal application-specific chemical compositions. Here we use an example of melting temperature (T-m) prediction to demonstrate the promise of machine learning (ML)-based techniques for establishing efficient structure-property mappings in PHA-based chemical space. We employ a manually curated data set of experimentally measured T-m values for a wide range of PHA homo- and copolymer chemistries along with their reported polymer molecular weights and polydispersity indices. Descriptors based on topology, shape, and charge/polarity of specific motifs forming the polymer backbone were then used to numerically represent the polymers. The ML models developed by using available data were used to rapidly predict the property of multicomponent PHA-based copolymers, while estimating uncertainties underlying the predictions. Combined with a previously developed glass transition temperature (T-g) prediction model and an evolutionary algorithm-based search strategy, the approach is demonstrated to address polymer design with multiobjective optimization challenges.

Automated summary

This article discusses the potential of using machine learning techniques to establish efficient structure-property mappings in the chemical space of PHAs. An example of predicting melting temperature is used to demonstrate the promise of this approach.