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

Area of Science:

• Polymer Chemistry
• Chemical Engineering
• Data Science

Background:

• Polymer physical properties are significantly influenced by the skew and shape of their molecular weight distribution (MWD).
• Traditional statistical summary metrics offer limited insight into the complete MWD.
• Predicting the entire MWD without information loss is crucial for material design.

Purpose of the Study:

• To demonstrate a high-throughput experimentation (HTE) platform for predicting polymer molecular weight distribution (MWD).
• To integrate machine learning (ML) models with HTE for comprehensive MWD prediction.
• To explore the application of transfer learning for efficient MWD prediction in batch polymerizations.

Main Methods:

• Developed a computer-controlled HTE system for parallel free radical polymerization of styrene under 8 variable conditions.
• Utilized inline Raman spectroscopy for real-time monomer conversion and offline size exclusion chromatography (SEC) for MWD analysis.
• Employed ML forward models to predict monomer conversion and MWD, incorporating SHAP analysis for interpretability.

Main Results:

• ML models accurately predicted monomer conversion, capturing varying polymerization kinetics across different experimental conditions.
• Entire MWDs, including skew and shape, were successfully predicted, with SHAP analysis revealing dependencies on reaction parameters.
• Transfer learning enabled accurate prediction of batch polymerization MWDs using minimal additional data points.

Conclusions:

• The synergistic combination of HTE and ML offers high predictive accuracy for polymerization outcomes.
• This integrated approach overcomes limitations of traditional MWD analysis.
• Transfer learning facilitates efficient exploration of synthesis parameter spaces, enabling targeted polymer design.