Interpretable Machine Learning Models for Phase Prediction in Polymerization-Induced Self-Assembly

Area of Science:

• Polymer Chemistry
• Materials Science
• Computational Chemistry

Background:

• Polymerization-induced self-assembly (PISA) is crucial for creating amphiphilic block copolymer structures.
• Predicting PISA phase behavior and morphologies is experimentally intensive, requiring empirical phase diagrams for new monomer pairs.
• Current methods lack efficiency for designing self-assemblies for specific applications.

Purpose of the Study:

• To develop the first data-driven framework for probabilistic modeling of PISA morphologies.
• To utilize machine learning methods to predict self-assembly behavior and reduce experimental burden.
• To create a tool that aids in designing empirical phase diagrams for novel monomers.

Main Methods:

• Curated a dataset of 592 data points from the PISA literature.
• Applied and adapted statistical machine learning methods, focusing on interpretable, low-variance models.
• Evaluated linear models, generalized additive models, and rule/tree ensembles for predictive performance.

Main Results:

• Machine learning models, excluding linear ones, showed decent interpolation performance (approx. 0.2 error rate) for known monomer pairs.
• The best model (random forest) demonstrated nontrivial extrapolation performance (0.27 error rate) for new monomer combinations.
• Active learning using the model efficiently guided phase diagram creation with minimal experiments (5-16 data points).

Conclusions:

• The developed data-driven framework effectively models PISA morphologies.
• Machine learning, particularly random forest, can significantly accelerate the creation of empirical phase diagrams.
• Publicly available data and code facilitate further research and application of this methodology.