(577e) Machine Learning for the Prediction of Glass Transition Temperature of Biodegradable Polymers | AIChE

(577e) Machine Learning for the Prediction of Glass Transition Temperature of Biodegradable Polymers


Kalaga, E. A. - Presenter, South Dakota School of Mines and Technology
Brenza, T., South Dakota Mines
The glass transition temperature (Tg) is the temperature at which the amorphous polymer changes its state from glassy to rubbery. Tg depends on the structural features of the polymers. These structural features include cross-linking in the polymer structure, polymer chain stiffness, molecular weight, number of carbon atoms etc. In general, chemical groups and intermolecular forces that are present in the polymer structures are the main factors that influence the Tg of polymers. This specific information about the structural features can be numerically encoded as molecular descriptors, a unique value associated with the chemical constitution for correlation of chemical structure with various physical properties. The purpose of the quantitative structure – property relationship (QSPR) studies is to find the relationship between the structure and the property of interest using these molecular descriptors.

In this study, we are using machine learning models such as multiple linear regression (MLR), support vector machine-based regression, random forest regression, K- nearest neighbors regression (KNN) and artificial neural networks (ANN) to construct a QSPR model. The QSPR model will be applied to the prediction of Tg for 59 biodegradable polymers having low Tg (from -35ºC to 225 ºC), 85 synthetic nondegradable polymers with moderate Tg (from 250 ºC to 399 ºC), and 77 aromatic polymers with high Tg (from 500 ºC to 800 ºC). The chemical structures of all the polymers were drawn in ChemDraw 2D and exported to ChemDraw 3D. The 3D structures were then optimized by applying energy minimization using molecular mechanics (MM2) until the rms gradient value became smaller than 0.1 kcal/mol Å. The optimized structures were further optimized using MOPAC method until rms value was reduced to smaller than 0.0001 kcal/mol Å. 1825 molecular descriptors were generated from the optimized chemical structures using the algorithms in the Mordred web graphical user interface (GUI) of python. Out of 1825, six descriptors were selected for machine learning models by means of stepwise regression followed by least absolute shrinkage and selection operator (Lasso) regularization methods using MATLAB. Using these six features the proposed machine learning models are being trained on 80 percent of the dataset created using Mordred GUI and tested on remaining 20 percent of the dataset for the prediction of Tg. We are also investigating the classification techniques such as decision trees, random forests, and support vector machine classifier (SVC) to classify polymers into different classes based on the relation between the selected molecular descriptors and Tg.

Keywords: Glass transition temperature, Biodegradable polymers, QSPR, Machine learning.