Prediction of Glass Transition Temperature of Polymers Using Simple Machine Learning

Area of Science:

• Materials Science and Engineering
• Computational Chemistry
• Polymer Science

Background:

• Polymer materials are crucial in various industries, with their glass transition temperature (Tg) being a key property for operational safety.
• Traditional Tg measurement methods (DSC, DMA) are accurate but often time-consuming, costly, and prone to errors.
• Developing efficient and reliable predictive models for polymer properties is essential for material design and application.

Purpose of the Study:

• To investigate the efficacy of Simplified Molecular Input Line Entry System (SMILES) as molecular descriptors for predicting polymer glass transition temperature (Tg).
• To compare the performance of various machine learning models (KNN, SVR, XGBoost, ANN, RNN) in predicting Tg using SMILES data.
• To evaluate different data conversion methods (One Hot Encoding and Natural Language Processing) for SMILES descriptors.

Main Methods:

• Utilized five machine learning models: k-nearest neighbors (KNN), support vector regression (SVR), extreme gradient boosting (XGBoost), artificial neural network (ANN), and recurrent neural network (RNN).
• Converted SMILES strings into numerical data using One Hot Encoding (OHE) and Natural Language Processing (NLP) techniques.
• Analyzed the impact of SMILES descriptor length on model performance, identifying optimal input ranges (200-200 characters).

Main Results:

• The Artificial Neural Network (ANN) model achieved the highest R² of 0.79, while the XGBoost (XGB) model demonstrated superior stability and faster training times with an R² of 0.774.
• One Hot Encoding (OHE) significantly reduced training times across most models compared to Natural Language Processing (NLP).
• The XGBoost model showed robustness in predicting new polymer data, with an average deviation of 9.76 from actual Tg values.

Conclusions:

• Simplified Molecular Input Line Entry System (SMILES) combined with machine learning offers an efficient alternative for predicting polymer glass transition temperature (Tg).
• The XGBoost model, utilizing One Hot Encoding for SMILES conversion, is recommended for Tg prediction due to its balance of accuracy, stability, and computational efficiency.
• Optimizing SMILES conversion strategies and model parameters is crucial for enhancing prediction reliability and advancing polymer material design.