Using Active Learning for the Computational Design of Polymer Molecular Weight Distributions

The design of the reaction conditions is essential for controlling polymerization to synthesize polymers with desired properties. However, the experimental screening of the reaction conditions can be time-consuming and costly. Computational methods such as kinetic Monte Carlo (KMC) simulations have been developed to assist with the design of experiments. Nevertheless, when the dimensions of the reaction conditions to be explored are large, the simulation models might still not be able to meet the demand for efficient screening and design. Active learning can be used to tackle this problem by designing data acquisition strategies that can minimize the labeling required to construct a good surrogate model in the design space. In this work, we combined a cluster-maximum model change (CMMC) model with KMC simulations, which can precisely design polymerization conditions at the lowest computational cost for the desired molecular weight distributions. The case study results show that the CMMC model only uses 50 KMC simulations to construct a predictive model with a 10% relative error for a system with 4 design parameters, which greatly reduces the computational cost while maintaining accuracy.