Replacing conventional, petro-based materials with bio-based materials will contribute to reductions in green-house-gas (GHG) emissions necessary for limiting the worst impacts of climate change. New materials development requires extensive experimental efforts, and the design space available through bio-accessible monomers is large and cannot be reasonably probed using the classic Edisionian approach. To aid in the identification of materials which can outperform existing materials, thus provide a market pull for their adoption, high-throughput screening tools can reduce the number of experimental candidates prior to synthesis. In this work we developed such a tool, a message passing neural network (MPNN) capable of predicting 8 material properties, glass transition temperature, melt temperature, density, modulus, and the gas permeability of O₂, N₂, CO₂, and H₂O, was developed. By probing the relationship between network depth and structural representation of polymers beyond simple repeat units, it was shown that larger representations of the polymer structure (i.e. higher degree of polymerization) can improve network performance. A domain of validity method was developed to verify the predicted materials had sufficient overlap with the polymer training set, and approaches for gaining insight into quantitative-structure-property-relationships using end-to-end network frameworks were developed. The end-to-ending learning network was then applied to four metabolic databases to predict properties of 1.4×10⁶ bio-accessible polymer candidates. Predictions of glass transition were benchmarked against experimentally determined glass transition temperatures to demonstrate the ability to predict properties of polymers that have not been previously synthesized. The development of this machine learning-based, polymer prediction tool will increase the rate at which new, sustainable materials can be discovered.

Mark Nimlos

Natl Renewable Energy Lab

Appearances