Accelerating Copolymer Design via Machine Learning

Establishing correlations between the sequence of monomers in a macromolecule and its properties is a grand challenge in polymer science. The major bottleneck in understanding a copolymer's sequence-structure and sequence-property correlations is its astronomically large sequence space that is intractable via standard experiments and first-principle-based computations. Machine learning can address this problem and help rapidly predict sequence-defined properties of copolymers. It could aid in establishing universal sequence-structure-property correlations and their design for a target property. Here, we survey and assess the efficacy of several machine learning methods for predicting copolymer properties and propose methods to improve their performance and transferability. We further propose a new ML pipeline that can tackle both the forward (property prediction) and inverse (sequence prediction) problems. We implement these approaches for three representative cases, viz., the single-molecule 3D structure of a copolymer in a dilute solvent condition, copolymer compatibilizer, and adsorption isotherm of a copolymer, demonstrating the generality of the proposed strategies.