Improvement of Extrapolative Prediction Performance by Quantum Chemical Calculations for Machine Learning Predictions of Flow-Synthesized Binary Copolymers.

The properties of polymers are highly dependent on the combination and composition ratio of the monomers used to prepare them; however, the large number of available monomers makes an exhaustive investigation of all the possible combinations difficult. In the present study, five binary copolymers were prepared by radical polymerization using a flow reactor, and the prediction performance of a machine learning model was evaluated in the interpolation and extrapolation regions for the estimation of the monomer conversion and monomer composition ratio in the polymer, which were used as objective variables. The prediction model was constructed using the process variables during polymerization and additional molecular descriptors (i.e., molecular flags (one-hot encoding), fingerprints or quantum chemical calculation values) related to the monomer type as explanatory variables. In the interpolated regions where all monomer types used were included in the training data, the prediction accuracy was high irrespective of the molecular descriptors. In the extrapolation region, the model that included explanatory variables corresponding to quantum chemical calculation values representing the energy related to the radical reactions such as energies of transition states and radicals, showed a high prediction accuracy for each objective variable. We found that quantum chemical calculation are important factors in the search for new binary copolymers. Besides, we will make a presentation about Bayesian optimization to synthesize a desired copolymer by optimizing the process variables.