Leveraging molecular descriptors and explainable machine learning for monomer conversion prediction in photoinduced electron transfer-reversible addition-fragmentation chain transfer polymerization

TL;DR

This paper introduces a machine learning model that predicts monomer conversion in a specific type of polymerization process using molecular descriptors and provides interpretable insights into the factors affecting conversion.

Contribution

The novel approach decomposes polymer systems into individual components and uses molecular descriptors with explainable ML to predict and interpret monomer conversion.

Findings

CatBoost was identified as the top-performing ML algorithm with an R2 of 0.84 for predicting monomer conversion.

SHAP analysis showed that monomer topological complexity, electronic polarization, and molecular weight explain over 60% of the model's predictive power.

The model generalized well to unseen (meth)acrylates and (meth)acrylamides with a MAE of 8.03.

Abstract

This study presents a molecular descriptor-based machine learning (ML) architecture for predicting monomer conversion in photoinduced electron transfer-reversible addition-fragmentation chain transfer (PET-RAFT) polymerization systems. Unlike traditional polymer informatics approaches that treat polymers as single units or use one-hot encoding for reaction components, we decompose each PET-RAFT system into its individual parts: monomer, RAFT agent, and photocatalyst. Next, each element was separately encoded using 2D molecular descriptors derived from SMILES. Using a literature-sourced dataset of 152 PET-RAFT systems, we systematically trained (with fivefold cross-validation, CV) and evaluated 10 ML algorithms. CatBoost showed greater stability across CV-folds (SD = ± 0.07) and was identified as the top performer for monomer conversion prediction (R2 = 0.84; RMSE = 10.04 pps; MAE = 8.16 pps). SHapley Additive exPlanations (SHAP) analysis revealed mechanistically interpretable structure–property-performance relationships, highlighting that monomer topological complexity, electronic polarization, and molecular weight together account for over 60% of the model’s predictive power. External validation confirmed CatBoost’s ability to generalize to unseen (meth)acrylates and (meth)acrylamides (MAE = 8.03), with comparable performance to that of the training set. In practice, the learned descriptor-conversion mapping enables fast in silico screening and component ranking, highlighting actionable descriptor ranges and potentially accelerating design-build-test cycles for high-conversion PET-RAFT.

The online version contains supplementary material available at 10.1038/s41598-025-33553-y.

Figures