Transfer Learning of Full Molecular Weight Distributions via High-Throughput Computer-Controlled Polymerization

The skew and shape of the molecular weight distribution(MWD) ofpolymers have a significant impact on polymer physical properties.Standard summary metrics statistically derived from the MWD only providean incomplete picture of the polymer MWD. Machine learning (ML) methodscoupled with high-throughput experimentation (HTE) could potentiallyallow for the prediction of the entire polymer MWD without informationloss. In our work, we demonstrate a computer-controlled HTE platformthat is able to run up to 8 unique variable conditions in parallelfor the free radical polymerization of styrene. The segmented-flowHTE system was equipped with an inline Raman spectrometer and offlinesize exclusion chromatography (SEC) to obtain time-dependent conversionand MWD, respectively. Using ML forward models, we first predict monomerconversion, intrinsically learning varying polymerization kineticsthat change for each experimental condition. In addition, we predictentire MWDs including the skew and shape as well as SHAP analysisto interpret the dependence on reagent concentrations and reactiontime. We then used a transfer learning approach to use the data fromour high-throughput flow reactor to predict batch polymerization MWDswith only three additional data points. Overall, we demonstrate thatthe combination of HTE and ML provides a high level of predictiveaccuracy in determining polymerization outcomes. Transfer learningcan allow exploration outside existing parameter spaces efficiently,providing polymer chemists with the ability to target the synthesisof polymers with desired properties.