Descriptor and Graph-based Molecular Representations in Prediction of Copolymer Properties using Machine Learning

This dataset accompanies a study that investigates the use of machine learning (ML) approaches for predicting seven different physical properties of 140 binary copolymers. Two computational methods were employed: a random forest (RF) model based on molecular descriptors and a Graph Neural Network (GNN) using 2D polymer graphs. These methods were applied in both single- and multi-task settings to explore the strengths of each approach in capturing various polymer properties.

The dataset includes two files:

Dataset.xlsx: Contains the following information for each of the 140 copolymers:

Polymer names

SMILES notation of the monomers

Fraction of monomers in each copolymer

Simulated values (calculated using  molecular dynamics simulation) and experimental values of various physical properties, including density, specific heat capacity at constant pressure ad volume ,radius of gyration, linear expansion coefficient, volume expansion coefficient, and bulk modulus .

Descriptors.xlsx: Provides the molecular descriptors calculated using PaDEL-Descriptor software, which were used as input for the RF model to predict polymer properties.

This work provides insight into the comparative strengths of descriptor- and graph-based representations in machine learning models for predicting material properties. It also highlights the importance of selecting appropriate molecular representations based on the nature of the properties being predicted.